A MIDAS Approach to Modeling First and Second Moment …...We propose a new approach to predictive...
Transcript of A MIDAS Approach to Modeling First and Second Moment …...We propose a new approach to predictive...
A MIDAS Approach to Modeling First and Second Moment
Dynamics∗
Davide PettenuzzoBrandeis University†
Allan TimmermannUCSD, CEPR, and CREATES‡
Rossen ValkanovUCSD§
April 24, 2015
Abstract
We propose a new approach to predictive density modeling that allows for MIDAS effectsin both the first and second moments of the outcome. Specifically, our modeling approachallows for MIDAS stochastic volatility dynamics, generalizing a large literature focusing onMIDAS effects in the conditional mean, and allows the models to be estimated by meansof standard Gibbs sampling methods. When applied to monthly time series on growth inindustrial production and inflation, we find strong evidence that the introduction of MI-DAS effects in the volatility equation leads to improved in-sample and out-of-sample densityforecasts. Our results also suggest that model combination schemes assign high weight toMIDAS-in-volatility models and produce consistent gains in out-of-sample predictive perfor-mance.
Key words: MIDAS regressions; Bayesian estimation; stochastic volatility; out-of-sampleforecasts; inflation forecasts, industrial production.
JEL classification: C53; C11; C32; E37
∗We thank three anonymous referees, Francesco Ravazzolo, Elena Andreou, and Eric Ghysels (the Editor) forvaluable comments and suggestions on an earlier version of the paper.†Brandeis University, Sachar International Center, 415 South St, Waltham MA 02453, Tel: (781) 736-2834.
Email: [email protected]‡University of California, San Diego, 9500 Gilman Drive, MC 0553, La Jolla CA 92093. Tel: (858) 534-0894.
Email: [email protected].§University of California, San Diego, 9500 Gilman Drive, MC 0553, La Jolla CA 92093. Tel: (858) 534-0898.
Email: [email protected].
1
1 Introduction
The notion that financial variables measured at a high frequency (e.g., daily interest rates
and stock returns) can be used to improve forecasts of less frequently observed monthly or
quarterly macroeconomic variables is appealing and has generated considerable academic interest
in a rapidly expanding literature on mixed-data sampling (MIDAS) models. MIDAS models
aggregate data sampled at different frequencies in a manner that has the potential to improve
the predictive accuracy of regression models. By using tightly parameterized lag polynomials
that allow the weighting on current and past values of the predictors to be flexibly tailored
to the data, the MIDAS approach makes it feasible to include information on long histories of
variables observed at a higher frequency than the outcome variable of interest.1
Empirical studies in the MIDAS literature have analyzed the dynamics in variables as diverse
as GDP growth (Andreou et al. (2013), Carriero et al. (2015), Clements and Galvao (2008),
Clements and Galvao (2009), Kuzin et al. (2011), Kuzin et al. (2013), Marcellino et al. (2013)),
stock market volatility (Ghysels et al. (2007), Ghysels and Valkanov (2012)) and the relation
between stock market volatility and macroeconomic activity (Engle et al. (2013) and Schorfheide
et al. (2014)). Such studies typically use simple and compelling designs to introduce high
frequency variables in the conditional mean equation and frequently find that the resulting
point forecasts produce lower out-of-sample root mean square forecast errors (RMSFEs) than
benchmarks ignoring high frequency information.
This paper introduces MIDAS-in-volatility effects using a Bayesian modeling approach that
allows for stochastic volatility (SV) dynamics and thus treats the underlying volatility as an
unobserved process. Previous studies either estimate MIDAS-in-mean models and allow for
SV effects (but not MIDAS-in-volatility) or estimate a MIDAS model on an observable proxy
for the volatility such as the realized variance. The premise of our approach is that there are
good reasons to expect information in variables observed at a high frequency to be helpful in
predicting the volatility of monthly or quarterly macroeconomic variables. Studies such as Sims
and Zha (2006) and Stock and Watson (2002) document that the volatility of macroeconomic
1A common alternative is to use an average of recent values, e.g., daily values within a quarter. However, thisoverlooks that recent observations carry information deemed more relevant than older observations. Alternatively,one can use only the most recent daily observation. However, this may be suboptimal, particularly in the presenceof measurement errors.
2
variables varies over time. Moreover, accounting for dynamics in the volatility equation can lead
to more efficient estimators and improved point forecasts as shown by Clark (2011), Carriero
et al. (2012), Carriero et al. (2015), and Clark and Ravazzolo (2014).2
The Bayesian modeling approach offers several advantages in our setting. Notably, the
objective of the Bayesian analysis is to obtain the predictive density given the data, as opposed
to simply computing a point forecast. Such density forecasts can be used to evaluate a range
of measures of predictive accuracy, including RMSFEs, log scores, and the continuously ranked
probability score of Gneiting and Raftery (2007a). Measures of the accuracy of density forecasts
are more likely to have power to detect improvements in volatility forecasts than measures
based on point forecasts such as the RMSFE. Moreover, our forecasts account for parameter
estimation error. This can be important in empirical applications with macroeconomic variables
for which data samples are short and parameters tend to be imprecisely estimated. Finally,
because we construct the predictive density for a range of different models, we can compute
forecast combinations that optimally weighs the individual models. Such forecast combinations
provide a way to deal with model uncertainty since they do not depend on identifying a single
best model. As new data arrive, the combination weights get updated and models that start
to perform better receive greater weight in the combinations.
Our paper exploits that the MIDAS lag polynomial can be cast as a linear regression model
with transformed daily predictors. For models with constant volatility and normal innovations,
Bayesian estimation can therefore be undertaken using a two-block Gibbs sampler. For models
with stochastic volatility we propose a specification that adds a MIDAS term to the log con-
ditional volatility equation. Conditional on the sequence of log-volatilities and the parameters
determining the stochastic volatility dynamics, our MIDAS specification reduces to a standard
linear regression model. In turn, to obtain the sequence of log-volatilities and the stochastic
volatility parameters we rely on the algorithm of Kim et al. (1998), extended by Chib et al.
2Carriero et al. (2015) develop a Bayesian method for producing current-quarter forecasts of GDP growththat is closely related to the U-MIDAS approach proposed by Foroni et al. (2013), and allow for both time-varing coefficients and stochastic volatility in the estimation. Ghysels (2012) extends the standard Bayesian VARapproach to allow for mixed frequency lags and MIDAS polynomials. Marcellino et al. (2013) develop a mixedfrequency dynamic factor model featuring stochastic shifts in the volatility of both the latent common factor andthe idiosyncratic components. Rodriguez and Puggioni (2010) cast a MIDAS regression model as a dynamic linearmodel, leaving unrestricted the coefficients on all the high frequency data lags.
3
(2002) to allow for exogenous covariates in the volatility equation.3 Hence a four-block Gibbs
sampler can be used to produce posterior estimates for the model parameters. Our approach
is straightforward to implement and well-suited for generating sequences of recursively updated
density forecasts.
We illustrate our approach in empirical applications to U.S. inflation and growth in industrial
production, both of which have been extensively studied in the literature. We use daily observa-
tions on eight predictor variables, including interest rates, stock returns and the business cycle
measure proposed by Aruoba et al. (2009). Because we estimate both MIDAS-in-mean and
MIDAS-in-volatility models, we can compare the contributions of high-frequency information
towards predicting first and second moments.
We find few instances where the MIDAS-in-mean effects lead to systematic improvements in
the point or density forecasts. In contrast, we find that MIDAS-in-volatility effects lead to sig-
nificantly better density forecasts of inflation and industrial production growth both in-sample
and out-of-sample. This holds even for benchmarks that are tough to beat such as factor models
with stochastic volatility (non-MIDAS) dynamics. Our finding holds across multiple forecast
horizons ranging from one through twelve months and across most of the daily predictor vari-
ables. Moreover, we find that model combination schemes produce better density forecasts than
a range of benchmarks, suggesting that our results are robust to model uncertainty. Moreover,
the Geweke and Amisano (2010) optimal prediction pool approach places at least 50 percent
of the weight on MIDAS-in-volatility models, further underscoring their importance to out-of-
sample performance in forecasting inflation and industrial production growth.
The outline of the paper is as follows. Section 2 introduces the MIDAS methodology and ex-
tends the model to include stochastic volatility effects and, as a new contribution, MIDAS terms
in both the conditional mean and the log conditional volatility equation. Section 3 introduces
our Bayesian estimation approach and discusses how to generate draws from the predictive den-
sity using Gibbs sampling methods. Section 4 describes our empirical applications, while Section
5 covers different forecast combination schemes. Section 6 concludes.
3Posterior simulation of the whole path of stochastic volatilities under an arbitrary second moment MIDASlag polynomial would require the use of a particle filter.
4
2 MIDAS regression models
This section outlines how we generalize the conventional regression specification to account for
MIDAS effects in the volatility equation, while also allowing for stochastic volatility.
2.1 MIDAS Setup
Suppose we are interested in forecasting some variable yt+1 which is observed only at discrete
times t− 1, t, t+ 1, etc., while data on a predictor variable, x(m)t , are observed m times between
t − 1 and t. For example, yt+1 could be a monthly variable and x(m)t could be a daily variable.
In this case m = 22, assuming that the number of daily observations available within a month
is constant and equals 22.
Let H ≥ 1 be an (arbitrary) forecast horizon and suppose we use the direct forecasting
approach to generate multi-period forecasts by projecting the period τ+H outcome on informa-
tion known at time t. It is natural to consider using lagged values of x(m)t to forecast yt+1. We
denote such lags of x(m)t by x
(m)t−j/m, where the m superscript makes explicit the higher sampling
frequency of x(m)t relative to yt+1. To include such lags we could use a simple MIDAS model4
yτ+H = β0 + B(L1/m;θ
)x(m)τ + ετ+H , τ = 1, ..., t−H (1)
where
B(L1/m;θ
)=
K−1∑k=0
B (k;θ)Lk/m.
Lk/m is a lag operator such that L1/mx(m)τ = x
(m)τ−1/m, and ετ+H is i.i.d. with E (ετ+H) = 0 and
V ar (ετ+H) = σ2ε . K is the maximum lag length for the included predictors. The distinguishing
feature of MIDAS models is that the lag coefficients in B (k;θ) are parameterized as a function of
a low dimensional vector of parameters θ = (θ0, θ1, ..., θp). To use a concrete example, suppose
again that yt+1 is a monthly series which gets affected by twelve months’ of daily data, x(m)t .
In this case, we would need K = 264 (22× 12) lags of daily variables. Without any restrictions
on the parameters in B(L1/m;θ
)there would be 264 + 2 parameters to estimate in (1). By
making B(L1/m;θ
)a function of a small set of p + 1 << K parameters we can greatly reduce
the number of parameters to estimate.
4For simplicity, our notation suppresses the dependence of the parameters on the forecast horizon, H.
5
It is sometimes useful to cast the MIDAS model as
yτ+H = β0 + β1B1
(L1/m;θ1
)x(m)τ + ετ+H , τ = 1, ..., t−H (2)
where β1B1
(L1/m;θ1
)= B
(L1/m;θ
)and β1 is a scalar that captures the overall impact of lagged
values of x(m)τ on yτ+H . Since β1 enters multiplicatively in (2), it cannot be identified without
imposing further restrictions on the polynomial B1
(L1/m;θ1
), e.g., by normalizing the function
B1
(L1/m;θ1
)to sum to unity.5
The model in (1) can be generalized to allow for py lags of yt and another pz lags of r
predictor variables zt = (z1t, ..., zrt)′ measured at the same frequency as yt :
yτ+H = α+
py−1∑j=0
ρj+1yτ−j +
pz−1∑j=0
γ ′j+1zτ−j + B(L1/m;θ
)x(m)τ + ετ+H . (3)
This regression requires the estimation of (2 + p+ py + pz × r) coefficients. The distributed lag
term∑py−1
j=0 ρj+1yτ−j captures same-frequency dynamics in yt+H , while the addition of the zt
factors allows for predictors other than own lags. We refer to the model in (3) as the Factor-
augmented AutoRegressive MIDAS, or FAR-MIDAS, for short. If the lagged factors are excluded
from equation (3), the model has only autoregressive and MIDAS elements and is called AR-
MIDAS. These abbreviations reflect the nested structure of the models that we consider.6
2.2 MIDAS weighting functions
The functional form of the MIDAS weights B(L1/m;θ
)depends on the application at hand and
has to be flexible enough to capture the dynamics in how the high frequency data x(m)τ affect
the outcome. We adopt a simple unrestricted version of B(L1/m;θ
), known as the Almon lag
polynomial, which takes the form
B (k;θ) =
p∑i=0
θiki, (4)
where θ = (θ0, θ1, ..., θp) is a vector featuring p + 1 parameters. Under this parameterization,
(3) takes the form
yτ+H = α+
py−1∑j=0
ρj+1yτ−j +
pz−1∑j=0
γ ′j+1zτ−j +
K−1∑k=0
p∑i=0
θikiLk/mx(m)
τ + ετ+H . (5)
5Normalization and identification of β1 are not strictly necessary in a MIDAS regression but can be usefulin settings such as those of Ghysels et al. (2005) and Ghysels et al. (2007) where β1 is important for economicinterpretation of the results.
6The FAR-MIDAS model is called FADL-MIDAS (for factor augmented distributed lag MIDAS) in Andreouet al. (2013).
6
Define the (p+ 1×K) matrix Q
Q =
1 1 1 . . . 11 2 3 . . . K1 22 32 . . . K2
......
.... . .
...1 2p 3p . . . Kp
, (6)
and the (K × 1) vector of high frequency data lags X(m)τ
X(m)τ =
[x(m)τ , x
(m)τ−1/m, x
(m)τ−2/m, ..., x
(m)τ−1, ..., x
(m)τ−(K−1)/m
]′. (7)
Given the linearity of (4) and (5), we can rewrite (5) as
yτ+H = α+
py−1∑j=0
ρj+1yτ−j +
qz−1∑j=0
γ ′j+1zt−j + θ′X(m)
τ + ετ+H , (8)
where X(m)
τ = QX(m)τ is a (p+ 1× 1) vector of transformed daily regressors. Once estimates
for the coefficients θ are available, we can compute the MIDAS weights from (4) as B (k;θ) =∑pi=0 θik
i. We can also impose the restriction that the weights B (k;θ) sum to one by normalizing
them as
B (k;θ) =B (k;θ)∑Ki=1 B (i;θ)
. (9)
In forecasting applications, this normalization does not provide any advantages. Hence, we
work with the unrestricted expression (8) for which the MIDAS parameters θ can conveniently
be estimated by OLS after transforming the daily regressors X(m)τ into X
(m)
τ .
It is useful to briefly contrast the Almon weights in (4) with other parameterizations in the
MIDAS literature. These include the exponential Almon lag of (Ghysels et al. (2005), Andreou
et al. (2013))
B (k;θ) =eθ1k+θ2k2+...+θpkp∑Ki=1 e
θ1i+θ2i2+...+θpip,
and the normalized Beta function of Ghysels et al. (2007)
B (k;θ) =
(k−1K−1
)θ1−1 (1− k−1
K−1
)θ2−1
∑Ki=1
(i−1K−1
)θ1−1 (1− i−1
K−1
)θ2−1.
Estimation of MIDAS models with these parameterizations requires non-linear optimization.
7
A third alternative is the stepwise weights proposed in Ghysels et al. (2007) and Corsi (2009)
B (k;θ) = θ1Ik∈[a0,a1] +
p∑p=2
θpIk∈[ap−1,ap],
where the p + 1 parameters a0, ..., ap are thresholds for the step function with a0 = 1 < a1 <
... < ap = K and Ik∈[ap−1,ap] is an indicator function, with
Ik∈[ap−1,ap] =
{1 if ap−1 ≤ k < ap0 otherwise
.
Provided that the threshold values are known, estimation of MIDAS models with these weights
can also be undertaken using OLS. A final alternative is the Unrestricted polynomial (U-MIDAS)
approach proposed by Foroni et al. (2013). In this case, all the high frequency lag coefficients
are left unconstrained, and estimation can be undertaken using OLS.
2.3 MIDAS in volatility
We next generalize the constant-volatility MIDAS models in the previous subsection to allow
for time-varying volatility and MIDAS-in-volatility effects. This generalization is potentially
important because it is well established that the use of high frequency variables leads to better
in-sample fit and out-of-sample forecasting performance for many financial and macroeconomic
variables; see Andersen et al. (2006), Ghysels et al. (2007), Engle et al. (2013), and Schorfheide
et al. (2014).
We generalize the constant volatility model in two steps. First, consider extending the FAR-
MIDAS model (8) to allow for stochastic volatility as
yτ+H = α+
py−1∑j=0
ρj+1yτ−j +
pz−1∑j=0
γ ′j+1zτ−j + θ′X(m)
τ + exp (hτ+H)uτ+H , (10)
where hτ+H denotes the log-volatility of yτ+H and uτ+H ∼ N (0, 1). It is commonly assumed
that the log-volatility evolves as a driftless random walk
hτ+H = hτ + ξτ+H , (11)
where ξτ+H ∼ N(
0, σ2ξ
)and ut and ξs are mutually independent for all t and s. We refer to
(10) and (11) as the FAR-MIDAS SV model. This type of specification is considered by Carriero
8
et al. (2015), and Marcellino et al. (2013), but with a different parameterization of the MIDAS
weights.7
The SV model in (11) permits time varying volatility but does not allow high frequency
variables, v(m)τ , to affect the conditional log-volatility. To accomplish this, we generalize (11) to
include second moment MIDAS effects
hτ+H = λ0 + λ1hτ +
K−1∑k=0
Bh (k;θh)Lk/mv(m)τ + ξτ+H . (12)
The daily variables v(m)τ need not be the same as those entering the first moment in (10), x
(m)τ .
Similarly, the MIDAS weights Bh (k;θh) need not be the same as those in the conditional mean
equation. The specification in (10) and (12) is a FAR-MIDAS with MIDAS stochastic volatility
or FAR-MIDAS SV-MIDAS model. In addition to allowing the high frequency lags to enter the
log-volatility equation, (12) relaxes the random walk assumption and introduces autoregressive
dynamics for the log-volatility.8 The stochastic volatility MIDAS specification is an analogue to
the MIDAS-in-mean specification, (3).
To complete the model in (12), we need to specify the SV-MIDAS weights, Bh (k;θh) , and
the v(m)τ variables. We focus on specifications for which the first moment and second moment
MIDAS variables are the same, x(m)τ = v
(m)τ , and apply Almon lag polynomials for both the first
and second moments. Under these assumptions, we can rewrite (12) as
hτ+H = λ0 + λ1hτ + θ′hX(m)
τ + ξτ+H , ξτ+H ∼ N(0, σ2
ξ
). (13)
We use the FAR-MIDAS SV-MIDAS model comprised of (10) and (13) in the estimation and
forecasting sections.
3 Bayesian estimation and forecasting
This section explains how we use Bayesian methods to estimate the MIDAS forecasting models
and generate density forecasts.
7The link between MIDAS models and time varying volatility has also been explored by Engle et al. (2013)who use a MIDAS-GARCH approach to link macroeconomic variables to the long-run component of volatility.Their model uses a mean reverting daily GARCH process and a MIDAS polynomial applied to monthly, quarterly,and biannual macroeconomic and financial variables.
8The addition of exogenous covariates in the log-volatility equation has been studied by Chib et al. (2002).
9
3.1 MIDAS models with constant volatility
Let Φ denote the regression parameters in the constant-volatility FAR-MIDAS model (3), ex-
cluding the MIDAS coefficients θ, i.e., Φ =(α, ρ1, ..., ρpy ,γ
′1, ...,γ
′qz
)′. Conditional on θ, the
MIDAS model reduces to a standard linear regression and one only needs to draw from the
posterior distributions of Φ and the variance of ετ+H , σ2ε . Assuming standard independent
Normal-inverted gamma priors on Φ and σ2ε and normally distributed residuals, ετ+H , drawing
from the posterior of these parameters is straightforward and simply requires using a two-block
Gibbs sampler.
The same logic extends to estimation of the MIDAS parameters θ in cases where the trans-
formed high frequency variables X(m)
t have a linear effect on the mean. Such cases include the
(non-normalized) Almon lag polynomial specification in (8), the step function polynomial spec-
ification of Ghysels et al. (2007), and the U-MIDAS approach of Foroni et al. (2013). Suppose
that ετ+H is normally distributed along with conjugate priors for the regression parameters
and error variance. For such cases a two-block Gibbs sampler can be used to obtain posterior
estimates for the parameters Φ, θ, and σ2ε .
9
To see how this works, rewrite (8) as
yτ+H = ZτΨ + ετ+H (14)
τ = 1, ..., t− 1
where Ψ=(Φ′,θ′
)′and Zτ =
(1, yτ , ..., yτ−py+1, z
′τ , ...,z
′τ−qz+1, X
(m)′τ
)′. Following standard
practice, suppose that the priors for the regression parameters Ψ in (14) are normally distributed
and independent of σ2ε
Ψ ∼ N (b,V ) . (15)
All elements of b are set to zero except for the value corresponding to ρ1 which is set to one.
Hence, our choice of the prior mean vector b reflects the view that the best prediction model is
9Under the U-MIDAS approach of Foroni et al. (2013), the matrix of transformed regressors, X(m)
t , is the
same as the original matrix, X(m)t .
10
a random walk. We choose a data-based prior for V :10
V = ψ2
s2y,t
(t−1∑τ=1
Z ′τZτ
)−1 , (16)
with
s2y,t =
1
t− 2
t−1∑τ=1
(yτ+1 − yτ )2 .
In (16), the scalar ψ controls the tightness of the prior. Letting ψ →∞ produces a diffuse prior
on Ψ. Our analysis sets ψ = 25, corresponding to relatively diffuse priors.
For the constant volatility model we assume a standard gamma prior for the error precision
of the return innovation, σ−2ε :
σ−2ε ∼ G
(s−2y,t , v0 (t− 1)
). (17)
v0 is a prior hyperparameter that controls how informative the prior is. v0 → 0 corresponds to
a diffuse prior on σ−2ε . Our baseline analysis sets v0 = 0.005. This corresponds to a pre-sample
of half of one percent of the actual data sample, again representing an uninformative prior.
LetDt denote information available at time t. Obtaining draws from the joint posterior distri-
bution p(Ψ, σ−2
ε
∣∣Dt) of the constant-volatility MIDAS regression model is now straightforward.
Combine the priors in (15)-(17) with the observed data to get the conditional posteriors:
Ψ|σ−2ε ,Dt ∼ N
(b, V
), (18)
and
σ−2ε
∣∣Ψ,Dt ∼ G(s−2, v
), (19)
where
V =
[V −1 + σ−2
ε
t−1∑τ=1
Z ′τZτ
]−1
,
b = V
[V −1b+ σ−2
ε
t−1∑τ=1
Z ′τyτ+1
], (20)
and
s2 =
∑t−1τ=1 (yτ+1 −ZτΨ)2 +
(s2y,t × v0 (t− 1)
)v
, (21)
10Priors for the hyperparameters are often based on sample estimates, see Stock and Watson (2006) and Efron(2010). Our analysis can be viewed as an empirical Bayes approach.
11
where v = v0 + (t− 1) .
For more general MIDAS lag polynomials, obtaining posterior estimates for θ is only slightly
more involved and requires a straightforward modification of the Gibbs sampler algorithm out-
lined above. As an example, Ghysels (2012) focuses on the case of normalized beta weights where
θ = (θ1, θ2)′, and suggests using a Gamma prior for both θ1 and θ2
θj ∼ G (f0, F0) , j = 1, 2. (22)
Here f0 and F0 are hyperparameters controlling the mean and degrees of freedom of the Gamma
distribution. Next, to draw from the posteriors of θ1 and θ2, Ghysels proposes utilizing a
Metropolis-in-Gibbs step as in Chib and Greenberg (1995). The Metropolis step is an accept-
reject step that requires a candidate θ∗ from a proposal density q(θ∗|θ[i]
), where θ[i] is the
last accepted draw for the MIDAS parameters θ. For example, when the Gamma distribution
is chosen for q(θ∗|θ[i]
), at iteration i+ 1 of the Gibbs sampler
θ∗j ∼ G(θ
[i]j , c
(θ
[i]j
)2)
, j = 1, 2, (23)
where c is a tuning parameter chosen to achieve a reasonable acceptance rate. The candidate
draw gets selected with probability min {a, 1} ,
θ[i+1] =
{θ∗ with probability min {a, 1}θ[i] with probability 1−min {a, 1} (24)
where a is computed as
a =L(Dt∣∣Φ,θ∗)
L(Dt|Φ,θ[i]
) G (θ∗| f0, F0)
G(θ[i]∣∣∣ f0, F0
) G(θ[i]∣∣∣θ∗, c (θ∗)2
)G(θ∗|θ[i], c
(θ[i])2) . (25)
L(Dt∣∣Φ,θ∗) and L
(Dt∣∣Φ,θ[i]
)are the conditional likelihood functions given the parameters
Φ,θ∗ and Φ, θ[i], respectively.
3.2 MIDAS models with time-varying volatility
Next, consider estimation of the models that allow the volatility of ετ+1 to change over time, as
in either (11) or (13). We focus our discussion on the most general process for the log-volatility,
(13), and note that when working with (11), λ0, λ1, and θh drop out of the model. For the FAR-
MIDAS SV-MIDAS model in equations (10) and (13), we require posterior estimates for all mean
12
parameters in equation (10), Ψ=(Φ′,θ′
)′, the sequence of log volatilities, ht = {h1, h2, ..., ht},
the parameters (λ0, λ1,θh), and the log-volatility variance σ2ξ .
We follow the earlier choice of priors for the parameters in the mean equation Ψ
Ψ ∼ N (b,V ) . (26)
Turning to the sequence of log-volatilities, ht = (h1, ..., ht), the error precision, σ−2ξ , and the
volatility parameters λ0, λ1, and θh, we can write
p(ht, λ0, λ1,θh, σ
−2ξ
)= p
(ht∣∣λ0, λ1,θh, σ
−2ξ
)p (λ0, λ1,θh) p
(σ−2ξ
).
Using (13), we can express p(ht∣∣λ0, λ1,θh, σ
−2ξ
)as
p(ht∣∣λ0, λ1,θh, σ
−2ξ
)=
t−1∏τ=1
p(hτ+1|λ0, λ1,θh, hτ , σ
−2ξ
)p (h1) , (27)
with hτ+1|λ0, λ1,θh, hτ , σ−2ξ ∼ N
(λ0 + λ1hτ + θ′hX
(m)
τ , σ2ξ
). To complete the prior elicitation
for p(ht, λ0, λ1,θh, σ
−2ξ
), we only need to specify priors for λ0, λ1, θh, the initial log-volatility,
h1, and σ−2ξ . We choose these from the normal-inverted gamma family
h1 ∼ N (ln (sy,t) , kh) , (28) λ0
λ1
θh
∼ N (mh,V h) , (29)
and
σ−2ξ ∼ G
(1/kξ, vξ (t− 1)
). (30)
We set kξ = 0.01, vξ = 1, and kh = 0.1. These are more informative priors than our earlier
choices. Setting kξ = 0.01 and vξ = 1 restricts changes to the log-volatility to be only 0.01 on
average. Conversely, kh = 0.1 places a relatively diffuse prior on the initial log volatility state.
We conduct a sensitivity analysis for these priors in a subsequent section.
Following Clark and Ravazzolo (2014) we set all the elements of the prior mean hyperpa-
rameter mh in (29) to zero, except for the parameter corresponding to the AR(1) coefficient
λ1, which we set to 0.9. As for the prior variance hyperparameter V h in (29), we set it to an
identity matrix with diagonal elements equal to 0.52, except for the element corresponding to
13
the AR(1) coefficient λ1, which we set to 0.012. This corresponds to a mildly uninformative
prior on the intercept and MIDAS coefficients, and a more informative prior on λ1, matching
the persistent dynamics in the log volatility process.
To obtain posterior estimates for the mean parameters Ψ, the sequence of log volatilities
ht, the stochastic volatility parameters (λ0, λ1,θh) , and the log-volatility variance σ2ξ , we use
a four-block Gibbs sampler to draw recursively from the following four conditional posterior
distributions:
1. p(
Ψ|ht, λ0, λ1,θh, σ−2ξ ,Dt
).
2. p(ht∣∣Ψ, λ0, λ1,θh, σ
−2ξ ,Dt
).
3. p(σ−2ξ
∣∣∣Ψ,ht, λ0, λ1,θh,Dt)
4. p(λ0, λ1,θh|Ψ,ht, σ−2
ξ ,Dt)
Simulating from the first three of these blocks is straightforward using the algorithms of Kim
et al. (1998), extended by Chib et al. (2002) to allow for exogenous covariates in the volatility
equation. The conditional posterior distribution of the SV-MIDAS parameters p(λ0, λ1,θh|Ψ,ht, σ−2
ξ ,Dt)
in the fourth step can be expressed as
λ0, λ1,θh|Ψ,ht, σ−2ξ ,Dt ∼ N
(mh,V h
),
where
V h =
V −1h + σ−2
ξ
t−1∑τ=1
1hτ
X(m)
τ
[1, hτ , X(m)′τ
]−1
, (31)
and
mh = V h
V −1h mh + σ−2
ξ
t−1∑τ=1
1hτ
X(m)
τ
hτ+1
. (32)
3.3 Forecasts from MIDAS models
The objective of Bayesian estimation of MIDAS forecasting models is to obtain the predictive
density for yt+1. This density conditions only on the data and so accounts for parameter un-
certainty. For example, working with the constant volatility MIDAS model (8), the predictive
14
density for yt+H is given by
p(yt+H | Dt
)=
∫Ψ,σ−2
ε
p(yt+H |Ψ,σ−2
ε ,Dt)p(Ψ,σ−2
ε
∣∣Dt) dΨdσ−2ε , (33)
where p(Ψ,σ−2
ε
∣∣Dt) denotes the joint posterior distribution of the MIDAS parameters condi-
tional on information available at time t, Dt.
Alternatively, when working with the FAR-MIDAS SV-MIDAS model in (10) and (13) the
density forecast for yt+H is given by
p(yt+H | Dt
)=
∫Ψ,ht+H ,λ0,λ1,θh,σ
−2ξ
p(yt+H |Ψ,ht+H ,ht, λ0, λ1,θh, σ
−2ξ ,Dt
)×p(ht+H |Ψ,ht, λ0, λ1,θh, σ
−2ξ ,Dt
)(34)
×p(
Ψ,ht, λ0, λ1,θh, σ−2ξ
∣∣∣Dt) dΨdht+Hdλ0dλ1dθhdσ−2ξ .
We can use the Gibbs sampler to draw from the predictive densities in (33) and (34). These
draws, y(j)t+H|t, j = 1, ..., J can be used to compute objects such as point forecasts, yt+H|t =
J−1∑J
j=1 y(j)t+H|t or the quantile of the realized value of the predicted variable, J−1
∑Jj=1 I(yt+H ≤
y(j)t+H|t), where I(yt+H ≤ y
(j)t+H|t) is an indicator function that equals one if the outcome, yt+H ,
falls below the jth draw from the Gibbs sampler.
3.4 Implementation of the Gibbs Sampler
We run the Gibbs samplers for 15,000 iterations, after a burn-in period of 1,000 iterations,
thinning the chains by keeping one out of every three draws.11 To evaluate convergence we
compute the following diagnostics: (1) autocorrelation functions of the draws (the smaller the
autocorrelations, the more efficient the samplers are); (2) inefficiency factors (IFs) for the pos-
terior estimates of the parameters. The IF is the inverse of the relative numerical efficiency
measure of Geweke (1992), i.e., the IF is an estimate of (1 + 2∑∞
k=1 ρk), where ρk is the k-th
autocorrelation of the chain. In our application the estimate is performed using a 4% tapered
window for the estimation of the spectral density at frequency zero. Values of the IFs below or
around 20 are regarded as satisfactory; (3) Raftery and Lewis (1992) diagnostics on the total
number of runs required to achieve a certain precision. The parameters for the diagnostics are
specified as follows: quantile = 0.025; desired accuracy = 0.025; required probability of attaining
11The calculations were performed on the High Performance Computing Cluster at Brandeis University.
15
the required accuracy = 0.95. The required number of runs should fall below the total number
of iterations used.
We compute these diagnostics for all estimated models. To preserve space we only report
results here for the FAR-MIDAS SV-MIDAS models, i.e., the most general of the models under
consideration.12 In all cases the 20th order sample autocorrelation of the draws is less than 0.06.
Similarly, in most cases the Geweke (1992) IF of the posterior parameter estimates is around
3, and is never larger than 7. Finally, the Raftery and Lewis (1992) diagnostic on the total
number of draws required is significantly below the total number of iterations used to estimate
our models. In summary, all convergence diagnostics appear satisfactory.
4 Empirical Results
This section introduces our monthly data on U.S. growth in industrial production and inflation
(our target variables), a set of macroeconomic factors, and the daily predictors. We then analyze
the in-sample and out-of-sample predictive accuracy of our forecasts for the model specifications
described in sections 2 and 3. We apply a range of measures to evaluate the predictive accuracy
of our forecasts. As discussed above, one of the advantages of adopting a Bayesian framework
is the ability to compute predictive distributions, rather than simple point forecasts, and to
account for parameter uncertainty. Accordingly, to shed light on the predictive ability of the
different models, we evaluate both point and density forecasts.
4.1 Data
Our empirical analysis uses monthly data on U.S. industrial production and the inflation rate.
Specifically, Let Iτ denote the monthly seasonally adjusted Industrial Production Index (IPI) at
time τ , obtained from the Federal Reserve of St. Louis FRED database, and define
yτ = 100× ln (Iτ/Iτ−1) . (35)
Similarly, the monthly inflation rate is obtained as
yτ = 1200× ln (Pτ/Pτ−1) , (36)
12These results are based on the specification that predicts the growth in industrial production for H = 1,using as daily predictor the federal funds rate. Similar results are obtained for other forecast horizons and dailypredictors, and for either growth in industrial production or inflation.
16
where Pτ denotes the U.S. monthly seasonally adjusted Consumer Price Index for All Urban
Consumers (All Items) at time τ , again downloaded from the Federal Reserve of St. Louis FRED
database.
The monthly predictor variables are an updated version of the 132 macroeconomic series used
in Ludvigson and Ng (2009) and extended by Jurado et al. (2014) to December 2011. The series
are selected to represent broad categories of macroeconomic quantities such as real output and
income, employment and hours, real retail, manufacturing and trade sales, consumer spending,
housing starts, inventories and inventory sales ratios, orders and unfilled orders, compensation
and labor costs, capacity utilization measures, price indexes, bond and stock market indexes,
and foreign exchange measures.13 We follow Stock and Watson (2012) and Andreou et al.
(2013) and extract two common factors (zτ ) from the 132 macroeconomic series using principal
components.
We consider eight daily series, x(m)τ , in this study: (i) the effective Federal Funds rate
(Ffr), first-differenced to eliminate any trends; (ii) the interest rate spread between the 10-
year government bond rate and the federal fund rate (Spr); (iii) value-weighted returns on US
stocks (Ret); (iv) returns on the portfolio of small minus big stocks considered by Fama and
French (1993) (Smb); (v) returns on the portfolio of high minus low book-to-market ratio stocks
studied by Fama and French (1993) (Hml); (vi) returns on a winner minus loser momentum
spread portfolio (Mom), (vii) the ADS daily business cycle variable of Aruoba et al. (2009);
and (viii) the default spread measured as the difference in the yield on portfolios of BAA and
AAA-rated corporate bonds.14 The interest rate series are from the Federal Reserve Bank of
St. Louis database FRED. Value-weighted stock return data are obtained from CRSP and
include dividends. Returns on the Smb, Hml, and Mom spread portfolios are downloaded from
Kenneth French’s data library.15 The ADS series are published by the Federal Reserve Bank of
Philadelphia. Our data sample spans the period from 1962:01 to 2011:12.
To test whether a better in-sample fit and out-of-sample forecasts can be obtained by includ-
ing the daily series in the forecasting model through MIDAS polynomials, we estimate several
13The data are available on Sydney Ludvigson’s website at http://www.econ.nyu.edu/user/ludvigsons/jlndata.zip14Using an international sample of data on ten countries, Liew and Vassalou (2000) find some evidence that
Hml and Smb are helpful in predicting future GDP growth.15We thank Kenneth French for making this data available on his website, at
http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data library.html
17
versions of the MIDAS specifications discussed above. These fall into one of three categories:
(i) MIDAS-in-mean with constant volatility (8); (ii) MIDAS-in-mean with stochastic volatility
(10 and 11); and (iii) MIDAS in both the mean and the volatility (10 and 13).
To explore the importance of MIDAS-in-mean effects, we compare AR-MIDAS and AR-
MIDAS SV models to the corresponding non-MIDAS models. We do the same for the FAR-
MIDAS and FAR-MIDAS SV models. Similarly, to assess the importance of MIDAS-in-volatility
effects, we compare AR-MIDAS SV-MIDAS and FAR-MIDAS SV-MIDAS models to AR-MIDAS
SV and FAR-MIDAS SV models, respectively. In total we consider six different MIDAS speci-
fications.
Every MIDAS specification is estimated with one of the eight daily variables, Ffr, Spr, Ret,
Smb, Hml, Mom, ADS, and Def, as defined above. Therefore, we have a total of 48 MIDAS
forecasting models. In addition, we estimate four benchmark non-MIDAS models: a purely
autoregressive model (AR); the same model with stochastic volatility (AR SV); a model that
includes AR terms and factors (FAR); and a model with factors and stochastic volatility (FAR
SV). In all cases our analysis assumes four AR lags of y (py = 4), two lags of the macro factors
(pz = 2) and twelve months of past daily observations.16
4.2 In-sample estimates and model comparisons
We first compare the fit of the different model specifications over the full sample, 1962:01-
2011:12. In a Bayesian setting, a natural approach to model selection is to compute the Bayes
factor, B1,0, of the null model M0 versus an alternative model, M1. The higher is the Bayes
factor, the higher are the posterior odds in favor of M1 against M0. We report two times the
natural log of the Bayes factors, 2 ln(B1,0). To interpret the strength of the evidence, we follow
studies such as Kass and Raftery (1995) and note that if 2 ln(B1,0) is below zero, the evidence
supports M0 over M1. For values of 2 ln(B1,0) between 0 and 2, there is “weak evidence” that
M1 is a more likely characterization of the data than M0. We view values of 2 ln(B1,0) between
2 and 6, 6 and 10, and higher than 10, as “some evidence,” “strong evidence,” and “very strong
16Our baseline results impose no restrictions on the coefficients of the four AR lags. We separately investigatedthe importance of this assumption by adding a stationarity restriction on the coefficients of the four AR lags,specifying a truncated normal prior for Ψ with support over the stationary region. We implemented this restrictionby augmenting our Gibbs sampler with an accept/reject step that removes all explosive draws from the posteriorof Ψ, and found the stationary restriction to have virtually no impact on the results.
18
evidence”, respectively, in support of M1 versus the null, M0.
Table 1 shows pairwise model comparisons based on the transformed Bayes factors, 2 ln(B1,0).
First consider the results for the growth in industrial production shown in the left columns.
Panels A and B display results for the AR-MIDAS and FAR-MIDAS models (8) relative to AR
or FAR models, respectively, while Panels C and D reports the same statistics for models with
SV dynamics in the second moment. Each of the MIDAS models is estimated by including a
single daily predictor at a time, as shown in separate rows. At the shortest horizons (H = 1, 3
months) only the MIDAS models based on the Ret or ADS variables lead to strong improvements
in forecasting performance relative to the corresponding benchmarks. These results carry over
to the models that allow for stochastic volatility dynamics; by and large, ADS appears to be
the only predictor that leads to better forecasts at the one-month horizon (H = 1).
Panels E and F in Table 1 report Bayes factors for specifications with MIDAS effects in the
volatility relative to non-MIDAS SV models, assuming that the mean already includes a MIDAS
term. In other words, we test different versions of (10) and (13) versus the model implied by
(10) and (11). We use the simpler SV volatility specification in our benchmark to reflect the
popularity of this approach in empirical work; see Cogley et al. (2005), D’Agostino et al. (2013),
Stock and Watson (2007), as well as the empirical evidence reported in Clark and Ravazzolo
(2014). Also, empirical tests show that the RMSE performance generated by the two SV models
in (13) and (11) is statistically indistinguishable.17
In general, for both the AR-MIDAS, SV-MIDAS and FAR-MIDAS, SV-MIDAS models the
evidence strongly supports adding MIDAS effects to the volatility specification for the vast
majority of comparisons. The evidence in favor of the IPI models with MIDAS in volatility
grows stronger, the longer the forecast horizon.
For the inflation rate models shown in the right columns of Table 1 the results mirror the
IPI findings, namely, there is spotty evidence of improvement in the forecasting models due to
the inclusion of MIDAS effects in the conditional mean (Panels A-D). In contrast, there is very
strong evidence that the models that include MIDAS in the volatility equation lead to superior
forecasts (Panels E-F). For the inflation series there is less evidence of systematic patterns in
17While the more general SV model (13) produces significantly better log probability scores than (11), thismodel is outperformed by the best SV-MIDAS models that include daily variables such as the ADS index.
19
the Bayes factors related to the forecast horizon compared with the data on growth in industrial
production.
In summary, the (in-sample) evidence from the Bayes factors suggests that there are no gains
in the performance of the prediction models from including MIDAS effects in the conditional
mean. In contrast, we find strong evidence of improvements as a result of including MIDAS
terms in the volatility equation. This holds across predictor variables and at both short and
long forecast horizons.
4.2.1 Sensitivity to Choice of Priors
Readers may be concerned that the Bayes factors are sensitive to our choice of priors, particularly
in the case of the volatility parameters which are chosen to be relatively informative. To address
this point, we conduct a sensitivity analysis that highlights the robustness of our results by
varying the key prior hyper-parameters, ψ, v0, and kξ over a wide range. In particular, ψ varies
between 100 and 2.5 (we set ψ = 25 in our baseline results); v0 varies between 0.001 and 0.1
(v0 = 0.005 in our baseline results), and kξ varies between 0.0001 and 1 (kξ = 0.01 in our
baseline results).
Table 2 reports Bayes factor results obtained under these alternative priors for the daily
predictors Ffr, Ret, and ADS. The baseline results only change in meaningful ways when we
change the prior on kξ which significantly alters the dynamics in the volatility equation. For the
other prior parameters, conclusions obtained under the other choices of priors are in line with
our baseline results.
4.2.2 MIDAS weighting of daily predictors
An alternative, and more traditional, way to investigate the importance of the MIDAS weighting
scheme applied to the daily predictor variables is to compare our MIDAS models with Almon
polynomials to a set of alternative models featuring a simple time average of the high frequency
data in either the mean and/or the volatility equation. The latter set of models are obtained by
restricting the Q matrix in Equation (6) to only its first row. Thus, we can test the importance of
the MIDAS weighting schemes in the mean and volatility by running an F-test on the coefficients
20
(θ1, θ2, ..., θp) for the mean and (θh,1, θh,2..., θh,p) for the volatility.18 The results of such F-tests,
reported in Table 3, suggest that the simpler weighting scheme in many cases is rejected against
the nesting MIDAS structure. Specifically, for the industrial production growth rate data we
find that the interest rate spread, stock returns, ADS and the default spread variables require
the more flexible MIDAS weighting schemes than the simple time average. While the evidence is
somewhat weaker for the inflation rate data, we still find strong evidence that the more flexible
weighting scheme is required for the models that include the interest rate spread, ADS, and
default spread data.
4.3 Out-of-sample forecasts
We next turn our attention to out-of-sample forecasting performance. To generate these, we
use the first twenty years of data as an initial training sample, i.e., we estimate our regression
models over the period 1962:01-1981:12 and use the resulting estimates to predict the outcome
in 1982:01. Next, we include 1982:01 in the estimation sample, which thus becomes 1962:01-
1982:01, and use the corresponding estimates to predict the outcome in 1982:02. We proceed
recursively in this fashion until the last observation in the sample, producing a time series of
one-step-ahead forecasts spanning the time period from 1982:01 to 2011:12.
4.3.1 Point forecasts
First consider the performance of the point forecasts. For each of the MIDAS models we obtain
point forecasts by repeatedly drawing from the predictive densities, p (yτ+H |Mi,Dτ ), and aver-
aging across draws. We have added Mi in the conditioning argument of the predictive density
to denote the specific model i, while τ ranges from 1981:12 to 2011:12-H. Following Stock and
Watson (2003) and Andreou et al. (2013), we measure the predictive performance of the MIDAS
models relative to a random walk (RW) model. Specifically, we summarize the precision of the
point forecasts of model i, relative to that from the RW model, by means of the ratio of RMSFE
values
RMSFEi =
√1
t−t+1
∑tτ=t e
2i,τ√
1t−t+1
∑tτ=t e
2RW,τ
, (37)
18We use the means and variance-covariance of the posterior distributions for (θ1, θ2, ..., θp) and (θh,1, θh,2..., θh,p)as inputs to these F tests.
21
where e2i,τ and e2
RW,τ are the squared forecast errors at time τ generated by model i and the RW
model, respectively. t and t denote the beginning and end of the forecast evaluation sample.
Values less than one for RMSFEi indicate that model i produces more accurate point forecasts
than the RW model.
The top panels in Figure 1 plot the sequence of recursively generated point forecasts of the
IPI growth rate generated by random walk, AR-MIDAS, AR-MIDAS SV, and AR-MIDAS SV-
MIDAS models fitted to short (H = 1) and long (H = 12) horizons. The RW forecasts are quite
different and notably more volatile than those from the three other models, particularly at the
longest horizon. A similar pattern holds for the inflation rate data plotted in the lower panels,
although the differences are attenuated for this series.
Figure 2 plots volatility forecasts for the same set of models. These are again generated
recursively and so are updated as new data arrives. This explains why the volatility forecast from
the AR-MIDAS model displays a slight downward trend, reflecting the lower average volatility
in the data following the start of the Great Moderation. The two SV models show considerable
short-term variation in the conditional volatility. However, the dynamics are markedly different.
For instance, the pure SV model generates more extreme volatility forecasts than the model
that includes MIDAS effects in the volatility equation, notably during 2007-2009.
Table 4 presents results for the RMSFE ratio in (37) using the IPI growth rate data. The
first column (marked “No MIDAS”) shows the RMSFE ratio for the benchmark models that do
not include MIDAS effects (i.e., AR, FAR, AR-SV, and FAR-SV models), while the subsequent
columns show RSMFE values for the different MIDAS models, each of which includes a single
daily predictor variable. We present results for different model specifications that add MIDAS
effects to the mean (rows 1, 2, 3, and 5) or to the volatility (rows 4 and 6) and across forecast
horizons ranging from H = 1 to H = 12 months.
An immediate observation from Table 4 is that all RMSFE ratios are well below one, typi-
cally ranging from 0.70 to 0.80, thus suggesting that the models reduce the RMSFE by 20-30%
relative to the random walk benchmark.19 At the short horizon (H = 1) 25% of this improve-
ment comes from adopting an AR specification rather than imposing a unit root, while another
19The models’ RMSFE values are all significantly lower than that of the random walk benchmark so we do notreport evidence of statistical significance in this table.
22
5% improvement comes from adding factors. Second, the results show that using the MIDAS
approach to add daily stock returns (Ret) to the benchmark AR or FAR models results in small
(up to 3%), but systematic reductions in the RMSFE for forecast horizons up to H = 6 months.
At the shortest one-month horizon, MIDAS models based on the ADS reduce the RMSFE ratio
quite substantially. However, consistent with our findings in Table 1, these improvements seem
limited to the shortest forecast horizon.
Table 5 shows that reductions of 10-25% in the RMSFE of the random walk model are
obtained for the inflation rate series. However, for this variable the evidence is much weaker
that adding MIDAS to the mean and/or volatility equation results in more accurate out-of-
sample point forecasts.20
To test more formally if the MIDAS forecasts are more accurate than various competitors,
Table 6 reports Diebold-Mariano p-values under the null that a given model has the same pre-
dictive ability as the RW benchmark. Each case provides a pairwise comparison of a model with
MIDAS-in-mean effects to the corresponding model without such MIDAS effects. Consistent
with the results in Table 4, only the ADS variable generates significant reductions in the MSFE
values at the shortest one-month horizon for the IPI growth rate series (Panel A).21
4.3.2 Density forecasts
One limitation of the RMSFE values reported above is that they fail to capture the richness of the
MIDAS models as they do not convey the full information in the density forecast p(yt+1|Mi,Dt
).
Indeed, comparing the plots of the point forecasts and the volatility forecasts in Figures 1 and
2, it is clear that there are much greater differences between the volatility forecasts generated
by the different models. Figure 3 shows that such differences give rise to very different density
forecasts at two points in time. The left panels show the densities for 1994:01. Compared to
AR-MIDAS models, SV dynamics compress the predictive density for this period. SV-MIDAS
has a similar, but weaker, effect preserving some of the greater uncertainty associated with the
constant volatility forecasts. The second snapshot, shown in the right panels in Figure 3, occurs
20Our empirical results are obtained using a limited set of daily predictor variables representing stock returns,interest rates and the daily business cycle index proposed by Aruoba et al. (2009). It is possible that furtherimprovements could be obtained from MIDAS-in-mean effects by using a richer set of daily variables such as thatproposed by Andreou et al. (2014).
21A comparison of MSE-values for the models with MIDAS-in-volatility effects to the models that omit sucheffects yielded very similar results and so are not reported here to preserve space.
23
in the middle of the global financial crisis (2008:12). Here the SV model generates a much higher
conditional volatility than the constant-volatility benchmark, with the SV-MIDAS model again
falling in the middle.
To measure the importance of such differences, we follow Amisano and Giacomini (2007),
Geweke and Amisano (2010), and Hall and Mitchell (2007) and compute the average log-score
differential (LSD)
LSDi =
t∑τ=t
(LSi,τ − LSRW,τ ) , (38)
where LSi,τ (LSRW,τ ) denotes the log-score of model i (RW), computed at time τ. Positive values
of LSDi indicate that model i produces more accurate density forecasts than the RW model
and the larger the value, the greater the improvements.
Table 7 and Table 8 report results for the log-score measure for the IPI growth rate series
(Table 7) and the inflation rate (Table 8). Once again the first column reports the value of the
LSD for a given benchmark model that does not include MIDAS effects (i.e., AR, FAR, AR-SV,
and FAR-SV models). Compared with the RMSFE measure, the LSD results identify more
cases with improvements relative to the various benchmarks, particularly for the models that
include MIDAS-in-volatility effects (rows 4 and 6 in each panel). This makes sense because the
LSD measure reflects improved volatility forecasts whereas the RMSFE measure ignores such
improvements.
To assess the statistical significance of these results, Table 9 reports Diebold-Mariano p-
values for tests of equal predictive ability using the log-scores. We report results only for the
models that add MIDAS effects to the volatility since adding MIDAS to the mean does not result
in significant improvements as we have seen. Adding MIDAS effects to the volatility equation
produces significantly higher values of the log-score measure than the corresponding benchmark
model without such effects. This holds for most of the predictive variables, across most forecast
horizons, and for both the IPI growth rate and inflation rate series.
To see how the log score differential evolves over time, we compute the cumulative log score
differential for model i versus the RW model
CLSDi,t =
t∑τ=t
(LSi,τ − LSRW,τ ) . (39)
24
Positive and increasing values of CLSDi,t suggest that model i produces more accurate density
forecasts than the RW model.
Figure 4 plots cumulative log score differentials for a benchmark AR SV model, an AR-
MIDAS SV and an AR-MIDAS SV-MIDAS model, in each case selecting the MIDAS variable
that is most informative. Top panels show results for the IPI growth rate while bottom panels
show results for the inflation rate. At the 1-month horizon (left panel), the model that incorpo-
rates MIDAS-in-mean effects is notably better than the pure AR model. At the twelve-month
horizon (right panel) there is less to distinguish between the MIDAS-in-mean and AR models,
both of which are dominated by the SV-MIDAS model. For the inflation rate data shown in the
bottom plots, there are no benefits from adding MIDAS-in-mean effects whereas adding MIDAS-
in-volatility effects leads to better forecasting performance at both the one- and twelve-month
horizons.
Following Gneiting and Raftery (2007b), Gneiting and Ranjan (2011) and Groen et al. (2013),
we also compute the average continuously ranked probability score differential (CRPSD) of model
i relative to the RW model22
CRPSDi =
1t−t+1
∑tτ=tCRPSi,τ
1t−t+1
∑tτ=tCRPSRW,τ
. (40)
CRPSi,τ (CRPSRW,τ ) measures the average distance between the empirical cumulative distri-
bution function (CDF) of yτ (which is simply a step function in yτ ), and the empirical CDF
associated with the predictive density of model i (RW). Values less than one for CRPSDi suggest
that model i performs better than the benchmark RW model.
To save space we do not report the CRPSD results in separate tables but instead summarize
our findings. For the IPI growth rate data the CRPSD results fall between those obtained
for the RMSFE and LSD measures: for all MIDAS models based on the ADS variable we
observe improvements in the forecasts at the shortest horizon (H = 1). In addition, we observe
improvements across all horizons for the MIDAS-in-volatility specifications that include the
ADS variable. Similar, if somewhat weaker, results hold for the MIDAS models based on daily
aggregate stock returns (Ret). For the inflation rate series the CRPSD results are similar to
22Gneiting and Raftery (2007b) explain how the CRPSD measure circumvents some of the problems of thelogarithmic score, most notably the fact that the latter does not reward values from the predictive density thatare close to, but different from, the realization.
25
the RMSFE ratio results reported in Table 5 and there is very little evidence of systematic
improvements in forecasting performance due to adding MIDAS effects in the mean or volatility
specifications.
Figures 5 and 6 illustrate how the ranking of IPI growth rate and inflation models varies
across the RMSFE, LSD, and CRPSD criteria. Focusing on Figure 5, as the two figures contain a
similar message, the left panel plots performance according to the RMSFE ratio on the horizontal
axis against performance according to the log-score criterion on the vertical axis. The figure
shows very clearly that the SV-MIDAS models perform better than the SV models which in
turn outperform the linear models on the log score criterion. In contrast, there is not much to
differentiate between the SV and SV-MIDAS models according to the RMSFE criterion. Thus,
the benefits from using MIDAS models to incorporate daily information show up more strongly
in the density forecast measures than in point forecasts. The right windows in Figure 5 show
a close relation between the ranking by the RMSFE and CRPSD criteria within each class of
models. Importantly, however, the CRPSD measure ranks the SV-MIDAS models better than
the SV models which in turn perform better than the linear models, consistent with the ranking
by the LSD criterion. Figure 6 conveys the same information for inflation models across the
RMSFE, LSD, and CRPSD criteria.
4.3.3 Real time data
One concern about the factor augmented model is that many of the underlying macro variables
are not observed in real time and are subject to data revisions. To address this point, we estimate
models that replace the macroeconomic factors with the monthly NAPM PMI series which is
available in real time. We find that the density forecasts implied by the resulting models are
virtually indistinguishable from those we obtain from the baseline FAR models, thus suggesting
that the role of data revisions in our setting is somewhat limited.23
23In particular, we find that for both the IPI growth and inflation rate series the RMSFE ratio obtainedfrom comparing a FAR model to an AR model augmented with the NAPM PMI series is always statisticallyindistinguishable from one. Similarly, the log score differential between the same two sets of models is in almostall cases statistically indistinguishable from zero, with the exception of the inflation series for H = 12, for whichwe obtain an LSD value of 0.012 with a p−value of 0.030.
26
5 Forecast combinations
Our analysis covers multiple specifications that differ in terms of the identity of the daily MIDAS
variables, inclusion of macro factors, as well as assumptions about volatility dynamics. In
practice a forecaster will not know which, if any, model produces the best forecasts and so
is confronted with model uncertainty. An attractive strategy in this situation is to combine
forecasts from multiple models, rather than attempting to select a single best model or use a
large model that nests all other specifications.
5.1 Forecast combination schemes
To see how model combination works in our setting, let Mi denote a specific model and suppose
we have N different models with predictive densities{p(yt+1|Mi,Dt
)}Ni=1
. Our starting point
is the equal-weighted pool (EWP) which assigns equal weights to each model Mi
p(yt+1| Dt
)=
1
N
N∑i=1
p(yt+1|Mi,Dt
). (41)
Next, consider the optimal predictive pool proposed by Geweke and Amisano (2011),
p(yt+1| Dt
)=
N∑i=1
w∗t,i × p(yt+1|Mi,Dt
). (42)
We use the past predictive performance of the N models to recursively determine the (N × 1)
vector of weights w∗t =[w∗t,1, ..., w
∗t,N
]. This requires determining w∗t by solving a maximization
problem using only information available at time t,
w∗t = arg maxwt
t−1∑τ=1
log
[N∑i=1
wit × Sτ+1,i
], (43)
subject to w∗t ∈ [0, 1]N . Sτ+1,i = exp (LSτ+1,i) is the exponential of the recursively computed
log-score for model i at time τ + 1. We follow Geweke and Amisano (2010) and compute the
marginal likelihoods by cumulating the predictive log scores of each model, after conditioning
on an initial warm-up estimation sample, as
Pr(Dt∣∣Mi
)= exp
t∑τ=t
LSτ.i
. (44)
27
As t→∞ the weights chosen according to (43) minimize the Kullback-Leibler distance between
the data generating process and the combined density. Finally, we consider Bayesian model
averaging (BMA) which weighs the individual models by their posterior probabilities
p(yt+1| Dt
)=
N∑i=1
Pr(Mi| Dt
)p(yt+1|Mi,Dt
). (45)
Here Pr(Mi| Dt
)denotes the posterior probability of model i, computed using all information
available at time t,
Pr(Mi| Dt
)=
Pr(Dt∣∣Mi
)Pr (Mi)∑N
j=1 Pr (Dt|Mj) Pr (Mj). (46)
Pr(Dt∣∣Mi
)and Pr (Mi) are the marginal likelihood and prior probability for model i, respec-
tively. We assume that all models are equally likely a priori and so set Pr (Mi) = 1/N .
The idea of using forecast combination methods in MIDAS regressions is also explored by
Andreou et al. (2013) in the context of predicting quarterly GDP growth by means of a host of
daily financial variables. We note, however, that Andreou et al. (2013) combine point forecasts
from the individual MIDAS models whereas our approach combines density forecasts.
5.2 Empirical results for forecast combinations
Panels A and B of Table 10 report out-of-sample results for the RMSFE ratio and log-score
differential using the three combination schemes discussed above and so are directly comparable
to the values shown in Tables 4-9. Panels C, D, E and F conduct Diebold-Mariano tests against
the AR model (Panels C and D) or the AR-SV model (Panels E and F). Results for the IPI
growth rate are shown in the top panels while the bottom panels show results for the inflation
rate data.
For the IPI growth rate series the RMSFE ratios in Panel A are comparable to, or sometimes
even better than, the best of the univariate models reported in Table 4. A similar finding holds
for the log-scores in Panel B which are comparable to the best models from Table 7. At the
shortest horizon (H = 1) the results are particularly strong for the BMA and optimal prediction
pools, whereas the results tend to favor the equal-weighting scheme at horizons H ≥ 3. This
suggests that estimation errors in obtaining the combination weights matter more at the longer
horizons.
28
Using the Diebold-Mariano statistic to formally test if the combined forecasts produce bet-
ter out-of-sample results than specific benchmarks, we find that all three combination schemes
generate significantly better performance at the one-month horizon than both an AR bench-
mark (Panels C and D) and an AR-SV benchmark (Panels E and F). The results are a little
weaker at the longer horizons, although the equal-weighted combination continues to signifi-
cantly outperform the two benchmarks across most of the horizons and for both performance
metrics.
Quite similar results hold for the inflation rate data. Again we find that the model combi-
nations generate significantly better forecasts than the random walk model. Moreover, all three
combinations produce better forecasts than the AR benchmark, particularly when measured
along the log-score metric. For horizons up to H = 6 months the BMA and optimal prediction
pool produce significantly better forecasts than the AR-SV model as measured by the average
log-score.
To gain insights into how adding MIDAS effects to the volatility specification helps improve
the forecasts, Figure 7 provides a recursive plot of the combination weights assigned to different
classes of models. Except for the earliest part of the post-1982 sample, the AR-MIDAS and
AR-MIDAS SV models receive very little weight. In the remainder of the sample the two
models with SV-MIDAS dynamics receive most of the probability weight. The FAR-MIDAS
SV-MIDAS model is particularly important for H = 1 whereas the AR-MIDAS SV-MIDAS
receives significant weight for H = 12. The FAR-MIDAS SV model receives considerable weight
around 2000 in the inflation rate forecasts, particularly at the one-month horizon. Looking at
the figure from another perspective, the optimal weights assigned to the SV-MIDAS models is 50
percent or higher for the IPI growth rate and inflation at both horizons after an initial learning
period. This underscores the superior out-of-sample predictive ability of these models.
Turning to the probability mass assigned to MIDAS models containing the eight different
daily predictor variables, Figure 8 shows that the ADS business cycle variable receives most of
the weight in the short-run IPI growth rate forecasts (top panels), whereas the Ffr, Spr, Ret and
Hml variables are more important at the 12-month horizon. Less clear-cut results are obtained
for the inflation rate data (bottom panels).
A key observation from these plots is that the weighting of the different models fluctuates
29
considerably over time. Del Negro et al. (2014) develop a dynamic version of the prediction pool
approach which they show works well for combinations of two models. Extending their approach
to more than two models is computationally demanding so we stop short of implementing their
approach empirically. However, it is worth noting that the evolution in the combination weights
shown in Figures 7 and 8 resembles the type of patterns found by Del Negro et al. (2014).
In summary, these results show clear advantages from using forecast combinations as a way to
deal with model uncertainty. The superior predictive ability of SV-MIDAS models is manifest in
large weights from an optimal prediction pool. Moreover, our results clearly show time variations
in the optimal weights assigned to different MIDAS predictor variables and different types of
model specifications. By recursively updating the model weights, the forecast combinations
adaptively adjust to changes and produce better out-of-sample forecasts.
6 Conclusions
We develop a Bayesian approach to estimation of forecasting models that allows for MIDAS
effects in both the first and second moments of the predicted series along with stochastic volatility
dynamics. Our SV-MIDAS approach is easy to implement using conventional Gibbs sampling
and generates predictive densities that only condition on the data available when the forecasts
are made.
Empirical applications to monthly growth in industrial production and the monthly inflation
rate in the U.S. show that the inclusion of MIDAS-in-volatility effects leads to significantly better
in-sample and out-of-sample density forecasts compared with similar specifications without such
MIDAS effects. We do not find similar improvements using root mean squared error measures
based on the accuracy of point forecasts, nor do we find that adding MIDAS effects to the mean
equation leads to systematically better forecasts. Importantly, our results hold across multiple
forecast horizons stretching from one month through twelve months, and continue to hold even
when a state-of-the-art model with factors in the mean and stochastic volatility dynamics is
used as the benchmark.
Our results are obtained using daily data on individual variables. Some variables, notably
market-wide stock returns and the business cycle index of Aruoba et al. (2009), work better than
others when used in the MIDAS specification. This implies that model uncertainty could be an
30
issue. To handle this, we investigate a variety of model combination approaches. Empirically we
find strong evidence that such combinations produce significantly better out-of-sample density
forecasts.
Our empirical results open up a promising new venue for improving on macroeconomic and
financial volatility forecasts using the MIDAS-in-volatility approach. Previous studies have
found that it is difficult to improve out-of-sample volatility forecasts of variables such as stock
returns relative to a model that simply uses lagged volatility as a predictor (e.g., Paye, 2012).
However, such studies have not explored the combination of stochastic volatility dynamics with
MIDAS-in-volatility effects as proposed here.
References
Amisano, G. and R. Giacomini (2007). Comparing density forecasts via weighted likelihood
ratio tests. Journal of Business & Economic Statistics 25 (2), 177–190.
Andersen, T. G., T. Bollerslev, P. F. Christoffersen, and F. X. Diebold (2006). Volatility and
correlation forecasting. Volume 1 of Handbook of Economic Forecasting, pp. 777 – 878. Elsevier.
Andreou, E., E. Ghysels, and A. Kourtellos (2013). Should macroeconomic forecasters use daily
financial data and how? Journal of Business & Economic Statistics 31 (2), 240–251.
Aruoba, S. B., F. X. Diebold, and C. Scotti (2009). Real-time measurement of business condi-
tions. Journal of Business & Economic Statistics 27 (4), 417–427.
Carriero, A., T. E. Clark, and M. Marcellino (2012). Common drifting volatility in large bayesian
vars. CEPR Discussion Papers 8894.
Carriero, A., T. E. Clark, and M. Marcellino (2015). Realtime nowcasting with a bayesian mixed
frequency model with stochastic volatility. Journal of the Royal Statistical Society: Series A
(Statistics in Society), n/a–n/a.
Chib, S. and E. Greenberg (1995). Understanding the metropolis-hastings algorithm. The
American Statistician 49 (4), 327–335.
31
Chib, S., F. Nardari, and N. Shephard (2002). Markov chain monte carlo methods for stochastic
volatility models. Journal of Econometrics 108 (2), 281 – 316.
Clark, T. E. (2011). Real-time density forecasts from bayesian vector autoregressions with
stochastic volatility. Journal of Business & Economic Statistics 29 (3), 327–341.
Clark, T. E. and F. Ravazzolo (2014). Macroeconomic forecasting performance under alternative
specifications of time-varying volatility. Journal of Applied Econometrics, in press.
Clements, M. P. and A. B. Galvao (2008). Macroeconomic forecasting with mixed-frequency
data. Journal of Business & Economic Statistics 26 (4), 546–554.
Clements, M. P. and A. B. A. Galvao (2009). Forecasting us output growth using leading
indicators: an appraisal using midas models. Journal of Applied Econometrics 24 (7), 1187–
1206.
Cogley, T., S. Morozov, and T. J. Sargent (2005). Bayesian fan charts for u.k. inflation: Fore-
casting and sources of uncertainty in an evolving monetary system. Journal of Economic
Dynamics and Control 29 (11), 1893 – 1925.
Corsi, F. (2009). A simple approximate long-memory model of realized volatility. Journal of
Financial Econometrics 7 (2), 174–196.
D’Agostino, A., L. Gambetti, and D. Giannone (2013). Macroeconomic forecasting and struc-
tural change. Journal of Applied Econometrics 28 (1), 82 – 101.
Del Negro, M., R. Hasegawa, and F. Schorfheide (2014). Dynamic prediction pools: an investiga-
tion of financial frictions and forecasting performance. Penn Institute for Economic Research
Working Paper 14-034.
Efron, B. (2010). Large-Scale Inference: Empirical Bayes Methods for Estimation, Testing, and
Prediction. Cambridge University Press.
Engle, R. F., E. Ghysels, and B. Sohn (2013). Stock market volatility and macroeconomic
fundamentals. The Review of Economics and Statistics 95 (3), 776–797.
32
Foroni, C., M. Marcellino, and C. Schumacher (2013). Unrestricted mixed data sampling (midas):
Midas regressions with unrestricted lag polynomials. Journal of the Royal Statistical Society:
Series A (Statistics in Society) 178 (1), 57 – 82.
Geweke, J. (1992). Evaluating the accuracy of sampling-based approaches to calculating poste-
rior moments. In J. Bernado, J. Berger, A. Dawid, and A. Smith (Eds.), Bayesian Statistics
4, pp. 169 – 193. Clarendon Press, Oxford, UK.
Geweke, J. and G. Amisano (2010). Comparing and evaluating bayesian predictive distributions
of asset returns. International Journal of Forecasting 26 (2), 216 – 230.
Geweke, J. and G. Amisano (2011). Optimal prediction pools. Journal of Econometrics 164 (1),
130 – 141.
Ghysels, E. (2012). Macroeconomics and the reality of mixed frequency data. Department of
Economics, University of North Carolina Working Paper.
Ghysels, E., P. Santa-Clara, and R. Valkanov (2005). There is a risk-return trade-off after all.
Journal of Financial Economics 76 (3), 509 – 548.
Ghysels, E., A. Sinko, and R. Valkanov (2007). Midas regressions: Further results and new
directions. Econometric Reviews 26 (1), 53–90.
Ghysels, E. and R. Valkanov (2012). Forecasting volatility with midas. In Handbook of Volatility
Models and Their Applications, pp. 383–401. John Wiley & Sons, Inc.
Gneiting, T. and A. E. Raftery (2007a). Strictly proper scoring rules, prediction, and estimation.
Journal of the American Statistical Association 102, 359–378.
Gneiting, T. and A. E. Raftery (2007b). Strictly proper scoring rules, prediction, and estimation.
Journal of the American Statistical Association 102 (477), 359–378.
Gneiting, T. and R. Ranjan (2011). Comparing density forecasts using threshold- and quantile-
weighted scoring rules. Journal of Business & Economic Statistics 29 (3), 411–422.
Groen, J., R. Paap, and F. Ravazzolo (2013). Real-Time Inflation Forecasting in a Changing
World. Journal of Business and Economic Stastistics 31(1), 29–44.
33
Hall, S. G. and J. Mitchell (2007). Combining density forecasts. International Journal of
Forecasting 23 (1), 1 – 13.
Jurado, K., S. C. Ludvigson, and S. Ng (2014). Meauring uncertainty. American Economic
Review , forthcoming.
Kass, R. E. and A. E. Raftery (1995). Bayes factors. Journal of the American Statistical
Association 90 (430), pp. 773–795.
Kim, S., N. Shephard, and S. Chib (1998). Stochastic volatility: Likelihood inference and
comparison with arch models. The Review of Economic Studies 65 (3), 361–393.
Kuzin, V., M. Marcellino, and C. Schumacher (2011). Midas vs. mixed-frequency var: Nowcast-
ing gdp in the euro area. International Journal of Forecasting 27 (2), 529 – 542.
Kuzin, V., M. Marcellino, and C. Schumacher (2013). Pooling versus model selection for now-
casting gdp with many predictors: Empirical evidence for six industrialized countries. Journal
of Applied Econometrics 28 (3), 392–411.
Liew, J. and M. Vassalou (2000). Can book-to-market, size and momentum be risk factors that
predict economic growth? Journal of Financial Economics 57 (2), 221 – 245.
Ludvigson, S. C. and S. Ng (2009). Macro factors in bond risk premia. Review of Financial
Studies 22 (12), 5027–5067.
Marcellino, M. G., M. Porqueddu, and F. Venditti (2013). Short-term gdp forecasting with
a mixed frequency dynamic factor model with stochastic volatility. Bank of Italy Temi di
Discussione (Working Paper) No. 896.
Paye, B. S. (2012). ‘Deja vol’: Predictive regressions for aggregate stock market volatility using
macroeconomic variables. Journal of Financial Economics 106 (3), 527 – 546.
Raftery, A. and S. Lewis (1992). How many iterations in the Gibbs sampler? In J. Bernado,
J. Berger, A. Dawid, and A. Smith (Eds.), Bayesian Statistics 4, pp. 763–773. Clarendon
Press, Oxford, UK.
34
Rodriguez, A. and G. Puggioni (2010). Mixed frequency models: Bayesian approaches to esti-
mation and prediction. International Journal of Forecasting 26 (2), 293 – 311.
Schorfheide, F., D. Song, and A. Yaron (2014). Identifying long-run risks: A Bayesian Mixed-
Frequency approach. Unpublished working paper, University of Pennsylvania, Philadelphia,
PA.
Sims, C. A. and T. Zha (2006). Were there regime switches in u.s. monetary policy? American
Economic Review 96 (1), 54–81.
Stock, J. H. and M. W. Watson (2002). Has the business cycle changed and why? In M. Gertler
and K. Rogoff (Eds.), NBER Macroeconomics Annual, Volume 17. MIT Press.
Stock, J. H. and M. W. Watson (2003). Forecasting output and inflation: The role of asset
prices. Journal of Economic Literature 41 (3), 788–829.
Stock, J. H. and M. W. Watson (2006). Forecasting with many predictors. Volume 1 of Handbook
of Economic Forecasting, pp. 515 – 554. Elsevier.
Stock, J. H. and M. W. Watson (2007). Why has u.s. inflation become harder to forecast?
Journal of Money, Credit and Banking 39, 3–33.
Stock, J. H. and M. W. Watson (2012). Generalized shrinkage methods for forecasting using
many predictors. Journal of Business and Economic Statistics 30 (4), 481–493.
35
Table 1. Bayes factors
Panel A: AR-MIDAS vs. AR
Daily seriesIPI growth rate Inflation rate
H = 1 H = 3 H = 6 H = 9 H = 12 H = 1 H = 3 H = 6 H = 9 H = 12
Ffr 5.035 0.702 -1.227 -1.527 -2.085 12.566 2.534 -1.453 -9.971 5.990Spr -6.053 -26.141 -17.451 -8.096 3.646 11.873 7.453 -6.287 -13.761 -2.861Ret 20.478 28.045 7.986 2.773 -1.942 -3.607 -5.921 -4.744 -5.305 -4.179Smb 4.966 -0.061 -4.637 -5.252 -4.767 -1.058 -1.250 -4.831 -5.519 -6.226Hml -3.231 -4.199 -4.267 -2.500 -0.393 -5.216 -1.999 -1.706 -3.561 -7.859Mom -0.705 7.688 -5.670 -11.231 -16.720 -1.653 -5.624 -6.606 -13.098 -6.170ADS 85.414 2.114 4.607 3.411 -2.384 -1.945 0.243 -6.489 -5.122 -3.329Def 6.435 -3.377 -6.891 -9.013 -15.840 -3.451 -8.375 -14.600 -17.409 -13.951
Panel B: FAR-MIDAS vs. FAR
Ffr -2.099 -6.896 -1.572 -1.103 -4.457 -1.999 -0.852 -1.987 -7.694 -4.891Spr -0.961 -7.869 -12.290 -12.635 -12.137 -1.907 1.029 -3.540 -8.521 -2.466Ret 3.747 7.341 1.402 -1.401 -15.852 0.324 -2.677 -6.320 -7.112 -5.014Smb -2.219 -5.505 -3.567 -5.353 -7.285 2.169 -1.577 -7.517 -6.575 -7.477Hml -3.477 -3.753 -5.265 -4.319 -2.975 -4.755 -1.518 -2.118 -2.393 -3.471Mom -3.274 -3.776 -11.284 -9.386 -9.294 -3.477 -3.570 -6.756 -9.167 -5.447ADS 34.634 -2.313 2.936 -0.631 -6.367 2.265 15.480 -10.444 -13.588 -11.300Def 0.993 -5.984 -8.698 -5.164 -8.037 -1.722 5.834 -10.001 -14.275 -10.424
Panel C: AR-MIDAS SV vs. AR SV
Ffr -4.202 -9.050 -19.678 -9.468 -28.776 -3.492 -10.224 -2.380 -3.028 -1.154Spr -19.166 -40.150 -28.134 -25.714 -6.167 -2.287 -4.322 -14.816 -39.427 -26.722Ret 9.271 -2.800 -2.844 -13.269 -9.567 -2.388 -5.919 -28.664 -1.237 -7.630Smb 1.561 -8.388 -12.046 -16.369 -17.933 -0.570 1.588 -2.780 4.508 -5.152Hml -9.638 -15.631 -31.106 -27.338 -22.168 -5.448 -5.963 -10.014 -2.291 -10.410Mom -4.287 -9.398 -39.450 -36.212 -30.285 -3.941 -1.296 -22.837 4.467 -2.078ADS 66.997 4.527 9.210 6.420 -7.445 -2.960 -7.430 -18.625 -9.220 -18.920Def 1.671 -11.992 -33.778 -20.861 -18.997 4.234 -2.469 -0.468 -3.536 -1.093
Panel D: FAR-MIDAS SV vs. FAR SV
Ffr -3.672 -8.625 -12.921 -3.371 -24.363 -7.179 -8.728 -2.815 12.864 -3.761Spr 1.160 -11.630 -10.596 -25.894 -12.209 -3.641 -7.215 -15.783 4.471 -10.397Ret -0.930 -1.142 -7.295 -7.828 -25.213 -5.080 -12.999 -29.673 6.462 -15.077Smb -5.187 -13.988 -9.639 -9.327 -17.292 0.170 -1.993 10.936 14.268 -8.200Hml -6.695 -15.436 -26.916 -25.179 -24.520 -7.328 -8.311 -3.893 5.241 -8.380Mom -8.668 -15.073 -28.388 -20.395 -28.185 -2.634 -3.499 -30.219 14.843 -1.001ADS 38.654 -11.055 0.162 3.597 -8.821 4.567 3.353 -7.000 -4.409 -13.625Def -4.720 -16.080 -33.443 -17.634 -11.370 10.607 6.908 3.056 4.640 -0.812
Panel E: AR-MIDAS SV-MIDAS vs. AR-MIDAS SV
Ffr 9.962 19.527 38.172 41.127 59.056 21.726 31.325 61.524 27.415 27.252Spr 0.553 -3.654 25.815 8.516 66.973 29.973 50.073 55.670 2.331 -6.704Ret 27.660 38.825 49.499 35.354 69.606 19.531 36.202 54.915 30.917 21.801Smb 19.832 25.954 43.312 30.447 62.111 20.785 36.453 60.542 34.179 23.220Hml 7.954 18.359 26.601 26.642 67.721 9.155 36.214 59.045 33.045 18.866Mom 13.922 31.694 23.702 13.589 43.405 12.493 39.514 47.950 28.720 27.367ADS 79.708 36.043 56.360 44.895 70.462 15.739 28.742 43.332 14.847 5.825Def 15.107 23.652 16.992 16.378 45.410 20.792 28.700 46.399 16.192 0.757
Panel F: FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV
Ffr 22.758 26.663 29.059 43.199 57.174 26.241 34.665 67.709 30.086 29.776Spr 30.299 32.302 30.451 14.262 58.036 33.900 41.432 62.925 13.746 16.824Ret 27.510 42.061 33.666 35.622 46.425 31.173 32.218 53.554 25.713 22.762Smb 24.115 28.187 32.785 32.607 60.811 33.460 33.349 66.882 34.285 27.958Hml 20.403 24.142 22.600 27.687 59.379 19.027 31.942 64.711 24.572 21.799Mom 22.759 29.314 11.173 18.235 49.743 24.889 39.622 52.387 23.861 32.933ADS 57.213 33.414 39.822 41.413 64.113 33.828 44.358 46.618 15.392 18.570Def 23.176 18.889 -8.501 25.032 57.831 37.629 47.702 59.418 24.439 26.984
This table reports pairwise model comparisons using twice the natural logarithm of the Bayes factor, 2 × (lnB1,0), where
B1,0 denotes the Bayes factor obtained from comparing model M1 vs. M0 B1,0 = Pr(Dt∣∣M1
)/Pr
(Dt∣∣M0
). Pairwise
model comparisons are listed in the first column, where the notation ‘AR’ refers to an autoregressive model, ‘AR-MIDAS’
refers to an augmented distributed lag MIDAS model, ‘FAR’ refers to a factor augmented autoregressive model, and ‘FAR-
MIDAS’ refers to a factor augmented distributed lag MIDAS model. The suffixes ‘SV’ and ‘SV-MIDAS’ denote models with
stochastic volatility and MIDAS stochastic volatility, respectively. Column headers denote the daily predictor used in the
MIDAS models, namely: the effective Federal Funds rate (Ffr), the interest rate spread between the 10-year government
bond rate and the federal funds rate (Spr), value-weighted stock returns (Ret), the SML portfolio return (Smb), the HML
portfolio return (Hml), the MOM portfolio return (Mom), the business cycle variable of Aruoba et al. (2009) (ADS), and
the default spread between BAA and AAA rated bonds. Kass and Raftery (1995) suggest interpreting the results as follows:
For 2 × (lnB1,0) less than 0, the evidence favors M0 over M1; for 2 × (lnB1,0) 0-2, 2-6, 6-10, and higher than 10, there
is “weak evidence”, “some evidence”, “strong evidence”, and “very strong evidence” in favor of M1 relative to M0. Bold
numbers denote weak or stronger evidence in favor of M1.
Table 2. Bayes factors, alternative prior choices
AR-MIDAS SV vs. AR SV
(a) ψ = 100, v0 = 0.001, kξ = 0.1 (b) ψ = 2.5, v0 = 0.1, kξ = 0.1
Daily series
IPI growth rate Inflation rate IPI growth rate Inflation rate
H = 1 H = 12 H = 1 H = 12 H = 1 H = 12 H = 1 H = 12
Ffr -0.819 -29.199 -0.335 -20.352 -4.692 -2.073 2.811 -0.043Ret 10.779 -11.589 14.137 -5.611 -2.725 -4.973 -0.842 0.319ADS 69.926 -7.525 69.888 -2.774 -5.395 -26.804 0.897 2.450
(c) ψ = 25, v0 = 0.005, kξ = 0.01 (d) ψ = 25, v0 = 0.005, kξ = 1
Ffr 2.120 -9.915 4.528 -14.575 10.306 1.964 8.730 -10.657Ret 24.437 -0.700 26.182 4.071 2.975 -1.891 -0.574 -10.838ADS 85.006 3.966 68.382 3.443 -0.186 -20.695 -5.031 -32.238
FAR-MIDAS SV vs. FAR SV
(a) ψ = 100, v0 = 0.001, kξ = 0.1 (b) ψ = 2.5, v0 = 0.1, kξ = 0.1
Ffr 1.726 -23.044 -4.766 -20.588 -6.127 -8.232 1.072 -4.551Ret 3.879 -27.529 5.099 -21.932 -2.543 -18.765 -2.188 -1.835ADS 42.650 -10.782 43.305 -9.361 2.625 -19.495 1.057 -1.304
(c) ψ = 25, v0 = 0.005, kξ = 0.01 (d) ψ = 25, v0 = 0.005, kξ = 1
Ffr 1.820 -6.595 -2.997 -7.441 -2.025 -6.236 -1.054 -4.355Ret 5.916 -19.761 17.027 -41.753 -0.607 -8.643 0.934 -24.196ADS 38.818 -8.247 38.300 -12.530 5.672 -21.775 7.647 -6.888
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV
(a) ψ = 100, v0 = 0.001, kξ = 0.1 (b) ψ = 2.5, v0 = 0.1, kξ = 0.1
Ffr 8.316 60.694 6.066 53.630 20.346 22.493 10.200 14.390Ret 24.646 67.923 25.115 59.494 22.267 16.764 9.643 13.185ADS 78.306 68.327 80.780 59.034 17.677 -6.731 10.482 10.441
(c) ψ = 25, v0 = 0.005, kξ = 0.01 (d) ψ = 25, v0 = 0.005, kξ = 1
Ffr -27.646 -12.119 -47.857 12.815 1.542 1.919 29.526 -439.053Ret -7.881 -10.743 38.489 41.670 -10.166 -6.641 15.838 15.640ADS 54.511 -6.395 81.267 55.592 -10.243 -19.511 23.591 -6.302
FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV
(a) ψ = 100, v0 = 0.001, kξ = 0.1 (b) ψ = 2.5, v0 = 0.1, kξ = 0.1
Ffr 20.182 53.282 23.200 53.929 27.107 29.778 10.369 19.705Ret 25.108 44.802 31.008 47.488 31.300 18.359 11.329 17.826ADS 59.082 61.700 63.051 60.297 32.583 10.270 12.650 14.961
(c) ψ = 25, v0 = 0.005, kξ = 0.01 (d) ψ = 25, v0 = 0.005, kξ = 1
Ffr -24.640 -8.949 -15.869 -95.626 4.465 4.132 -57.439 -195.555Ret -15.360 -23.657 38.760 -10.402 7.004 -4.765 31.490 1.388ADS 14.059 -9.982 38.815 19.207 11.788 -9.928 31.260 16.555
This table reports pairwise model comparisons under four different prior choices. The table entries are equal to twice the
natural logarithm of the Bayes factor, 2×(lnB1,0), where B1,0 denotes the Bayes factor obtained from comparing model M1
vs. M0 B1,0 = Pr(Dt∣∣M1
)/Pr
(Dt∣∣M0
). Pairwise model comparisons are listed in the first column, where the notation
‘AR’ refers to an autoregressive model, ‘AR-MIDAS’ refers to an augmented distributed lag MIDAS model, ‘FAR’ refers to
a factor augmented autoregressive model, and ‘FAR-MIDAS’ refers to a factor augmented distributed lag MIDAS model.
The suffixes ‘SV’ and ‘SV-MIDAS’ denote models with stochastic volatility and MIDAS stochastic volatility, respectively.
Column headers denote the daily predictor used in the MIDAS models, namely: the effective Federal Funds rate (Ffr),
the interest rate spread between the 10-year government bond rate and the federal funds rate (Spr), value-weighted stock
returns (Ret), the SML portfolio return (Smb), the HML portfolio return (Hml), the MOM portfolio return (Mom), the
business cycle variable of Aruoba et al. (2009) (ADS), and the default spread between BAA and AAA rated bonds. Kass
and Raftery (1995) suggest interpreting the results as follows: For 2 × (lnB1,0) less than 0, the evidence favors M0 over
M1; for 2× (lnB1,0) 0-2, 2-6, 6-10, and higher than 10, there is “weak evidence”, “some evidence”, “strong evidence”, and
“very strong evidence” in favor of M1 relative to M0. Bold numbers denote weak or stronger evidence in favor of M1.
Table 3. F test on the statistical significance of the MIDAS weighting schemes
Daily series
IPI growth rate Inflation rate
H = 1 H = 3 H = 6 H = 9 H = 12 H = 1 H = 3 H = 6 H = 9 H = 12
Ffr 0.010 0.773 0.420 0.297 0.784 0.170 0.490 0.770 0.995 0.932Spr 0.171 0.028 0.005 0.018 0.010 0.088 0.235 0.067 0.007 0.008Ret 0.402 0.003 0.008 0.003 0.434 0.228 0.157 0.310 0.453 0.872Smb 0.234 0.337 0.762 0.276 0.547 0.140 0.307 0.707 0.427 0.932Hml 0.803 0.594 0.128 0.101 0.062 0.830 0.364 0.673 0.393 0.496Mom 0.944 0.105 0.132 0.493 0.010 0.342 0.106 0.104 0.354 0.144ADS 0.000 0.000 0.000 0.187 0.722 0.021 0.000 0.029 0.842 0.061Def 0.003 0.066 0.064 0.031 0.093 0.037 0.144 0.155 0.196 0.063
This table reports the p-values of an F test that the MIDAS coefficients (θ1, ..., θp) and(θh,1, ..., θh,p
)are jointly equal to
zero. We use the posterior means and variance-covariances of the posterior distributions for (θ1, ..., θp) and(θh,1, ..., θh,p
)as inputs in the F tests. Bold numbers indicate significance at the 10 percent level.
38
Table 4. Out-of-sample forecast performance for IPI growth rate - RMSFE
ModelNo
Ffr Spr Ret Smb Hml Mom ADS DefMidas
H = 1
AR-MIDAS 0.750 0.750 0.769 0.728 0.747 0.754 0.747 0.669 0.741FAR-MIDAS 0.683 0.687 0.682 0.679 0.684 0.686 0.684 0.647 0.679AR-MIDAS SV 0.737 0.740 0.747 0.715 0.735 0.737 0.735 0.652 0.725AR-MIDAS SV-MIDAS 0.745 0.750 0.718 0.738 0.742 0.738 0.657 0.728FAR-MIDAS SV 0.675 0.681 0.673 0.668 0.678 0.675 0.678 0.635 0.672FAR-MIDAS SV-MIDAS 0.683 0.675 0.672 0.680 0.678 0.680 0.639 0.673
H = 3
AR-MIDAS 0.804 0.814 0.871 0.775 0.807 0.809 0.786 0.811 0.810FAR-MIDAS 0.804 0.810 0.809 0.791 0.809 0.807 0.799 0.802 0.810AR-MIDAS SV 0.804 0.820 0.837 0.774 0.807 0.805 0.800 0.794 0.811AR-MIDAS SV-MIDAS 0.821 0.838 0.776 0.810 0.809 0.795 0.797 0.813FAR-MIDAS SV 0.791 0.799 0.799 0.779 0.799 0.792 0.788 0.792 0.796FAR-MIDAS SV-MIDAS 0.805 0.801 0.785 0.806 0.801 0.794 0.796 0.803
H = 6
AR-MIDAS 0.789 0.813 0.839 0.784 0.794 0.792 0.794 0.779 0.801FAR-MIDAS 0.778 0.783 0.796 0.776 0.781 0.784 0.788 0.775 0.792AR-MIDAS SV 0.791 0.810 0.815 0.778 0.795 0.793 0.813 0.775 0.802AR-MIDAS SV-MIDAS 0.813 0.818 0.779 0.795 0.793 0.808 0.776 0.804FAR-MIDAS SV 0.763 0.774 0.776 0.760 0.765 0.765 0.776 0.761 0.774FAR-MIDAS SV-MIDAS 0.779 0.779 0.764 0.769 0.768 0.779 0.762 0.781
H = 9
AR-MIDAS 0.750 0.765 0.772 0.755 0.759 0.752 0.764 0.743 0.763FAR-MIDAS 0.749 0.757 0.767 0.752 0.757 0.753 0.759 0.750 0.753AR-MIDAS SV 0.755 0.769 0.768 0.752 0.759 0.758 0.784 0.747 0.765AR-MIDAS SV-MIDAS 0.773 0.776 0.753 0.761 0.756 0.775 0.749 0.764FAR-MIDAS SV 0.749 0.763 0.763 0.750 0.753 0.754 0.761 0.750 0.751FAR-MIDAS SV-MIDAS 0.765 0.769 0.749 0.754 0.751 0.760 0.751 0.752
H = 12
AR-MIDAS 0.703 0.730 0.719 0.705 0.708 0.702 0.735 0.705 0.732FAR-MIDAS 0.694 0.723 0.726 0.713 0.703 0.695 0.712 0.700 0.703AR-MIDAS SV 0.714 0.729 0.707 0.719 0.720 0.715 0.736 0.719 0.734AR-MIDAS SV-MIDAS 0.732 0.711 0.715 0.717 0.712 0.733 0.716 0.736FAR-MIDAS SV 0.699 0.720 0.709 0.712 0.704 0.703 0.713 0.701 0.704FAR-MIDAS SV-MIDAS 0.723 0.715 0.714 0.706 0.703 0.715 0.704 0.707
This table reports the ratio between the RMSFE of model i and the RMSFE of the Random Walk (RW) model, computedas
RMSFEi =
√1
t−t+1
∑tτ=t e
2i,τ√
1t−t+1
∑tτ=t e
2RW,τ
,
where e2i,τ and e2RW,τ are the squared forecast errors at time τ generated by model i and the RW model, respectively,
and i denotes any of the models described in section 3. Values less than one for RMSFEi indicate that model i produces
more accurate point forecasts than the RW model. The notation ‘AR’ refers to an autoregressive model, ‘AR-MIDAS’
refers to an augmented distributed lag MIDAS model, ‘FAR’ refers to a factor augmented autoregressive model, and ‘FAR-
MIDAS’ refers to a factor augmented distributed lag MIDAS model. The suffixes ‘SV’ and ‘SV-MIDAS’ denote models
with stochastic volatility and MIDAS stochastic volatility, respectively. For MIDAS models, the column headers denote
the daily predictor used in the regressions, namely: the effective Federal Funds rate (Ffr), the interest rate spread between
the 10-year government bond rate and the federal funds rate (Spr), value-weighted stock returns (Ret), the SML portfolio
return (Smb), the HML portfolio return (Hml), the MOM portfolio return (Mom), the business cycle variable of Aruoba
et al. (2009) (ADS), and the default spread between BAA and AAA rated bonds. All forecasts and forecast errors are
produced with recursive estimates of the models. The out-of-sample period starts in 1982:1 and ends in 2011:12. Bold
numbers indicate the lowest RMSFE across all predictors for a given model.
Table 5. Out-of-sample forecast performance for inflation rate - RMSFE
ModelNo
Ffr Spr Ret Smb Hml Mom ADS DefMidas
H = 1
AR-MIDAS 0.918 0.912 0.905 0.921 0.921 0.921 0.928 0.919 0.930FAR-MIDAS 0.918 0.920 0.914 0.916 0.922 0.921 0.925 0.916 0.918AR-MIDAS SV 0.909 0.907 0.905 0.910 0.913 0.911 0.923 0.910 0.913AR-MIDAS SV-MIDAS 0.905 0.901 0.912 0.913 0.913 0.922 0.909 0.917FAR-MIDAS SV 0.911 0.915 0.916 0.910 0.915 0.913 0.923 0.913 0.908FAR-MIDAS SV-MIDAS 0.914 0.910 0.910 0.914 0.914 0.921 0.908 0.908
H = 3
AR-MIDAS 0.817 0.825 0.810 0.824 0.822 0.819 0.825 0.821 0.836FAR-MIDAS 0.819 0.829 0.816 0.825 0.825 0.820 0.826 0.808 0.813AR-MIDAS SV 0.815 0.822 0.812 0.817 0.814 0.815 0.817 0.814 0.823AR-MIDAS SV-MIDAS 0.822 0.811 0.819 0.816 0.815 0.817 0.812 0.827FAR-MIDAS SV 0.815 0.827 0.813 0.817 0.813 0.816 0.818 0.800 0.806FAR-MIDAS SV-MIDAS 0.825 0.811 0.818 0.816 0.816 0.818 0.799 0.807
H = 6
AR-MIDAS 0.795 0.792 0.800 0.804 0.796 0.801 0.804 0.809 0.820FAR-MIDAS 0.802 0.803 0.809 0.813 0.806 0.806 0.812 0.813 0.816AR-MIDAS SV 0.798 0.795 0.804 0.812 0.799 0.799 0.803 0.814 0.795AR-MIDAS SV-MIDAS 0.786 0.791 0.795 0.788 0.794 0.796 0.796 0.799FAR-MIDAS SV 0.806 0.806 0.811 0.816 0.800 0.806 0.811 0.810 0.791FAR-MIDAS SV-MIDAS 0.790 0.796 0.798 0.790 0.797 0.800 0.794 0.791
H = 9
AR-MIDAS 0.808 0.822 0.832 0.821 0.814 0.816 0.821 0.828 0.842FAR-MIDAS 0.819 0.837 0.842 0.834 0.827 0.823 0.831 0.837 0.843AR-MIDAS SV 0.798 0.799 0.824 0.791 0.787 0.790 0.790 0.800 0.790AR-MIDAS SV-MIDAS 0.801 0.805 0.794 0.790 0.794 0.796 0.800 0.795FAR-MIDAS SV 0.818 0.804 0.811 0.800 0.788 0.793 0.798 0.805 0.788FAR-MIDAS SV-MIDAS 0.809 0.814 0.805 0.793 0.797 0.803 0.804 0.791
H = 12
AR-MIDAS 0.786 0.785 0.802 0.794 0.792 0.801 0.793 0.798 0.810FAR-MIDAS 0.782 0.796 0.798 0.787 0.788 0.790 0.790 0.797 0.797AR-MIDAS SV 0.749 0.750 0.754 0.749 0.747 0.756 0.746 0.760 0.755AR-MIDAS SV-MIDAS 0.757 0.765 0.755 0.753 0.765 0.752 0.759 0.760FAR-MIDAS SV 0.750 0.755 0.756 0.752 0.747 0.753 0.748 0.753 0.747FAR-MIDAS SV-MIDAS 0.762 0.764 0.759 0.754 0.762 0.755 0.756 0.754
This table reports the ratio between the RMSFE of model i and the RMSFE of the Random Walk (RW) model, computedas
RMSFEi =
√1
t−t+1
∑tτ=t e
2i,τ√
1t−t+1
∑tτ=t e
2RW,τ
,
where e2i,τ and e2RW,τ are the squared forecast errors at time τ generated by model i and the RW model, respectively,
and i denotes any of the models described in section 3. Values less than one for RMSFEi indicate that model i produces
more accurate point forecasts than the RW model. The notation ‘AR’ refers to an autoregressive model, ‘AR-MIDAS’
refers to an augmented distributed lag MIDAS model, ‘FAR’ refers to a factor augmented autoregressive model, and ‘FAR-
MIDAS’ refers to a factor augmented distributed lag MIDAS model. The suffixes ‘SV’ and ‘SV-MIDAS’ denote models
with stochastic volatility and MIDAS stochastic volatility, respectively. For MIDAS models, the column headers denote
the daily predictor used in the regressions, namely: the effective Federal Funds rate (Ffr), the interest rate spread between
the 10-year government bond rate and the federal funds rate (Spr), value-weighted stock returns (Ret), the SML portfolio
return (Smb), the HML portfolio return (Hml), the MOM portfolio return (Mom), the business cycle variable of Aruoba
et al. (2009) (ADS), and the default spread between BAA and AAA rated bonds. All forecasts and forecast errors are
produced with recursive estimates of the models. The out-of-sample period starts in 1982:1 and ends in 2011:12. Bold
numbers indicate the lowest RMSFE across all predictors for a given model.
Table 6. Diebold-Mariano p-values from tests of equal predictive ability - RMSFE
Comparison Ffr Spr Ret Smb Hml Mom ADS Def
Panel A: IPI growth rate
H = 1
AR-MIDAS vs. AR 0.353 0.969 0.018 0.236 0.846 0.210 0.013 0.264AR-MIDAS SV vs. AR SV 0.509 0.945 0.012 0.326 0.631 0.220 0.003 0.133FAR-MIDAS vs. FAR 0.704 0.384 0.260 0.847 0.777 0.691 0.015 0.285FAR-MIDAS SV vs. FAR SV 0.871 0.300 0.146 0.857 0.546 0.808 0.004 0.299
H = 3
AR-MIDAS vs. AR 0.782 0.999 0.104 0.689 0.762 0.154 0.777 0.610AR-MIDAS SV vs. AR SV 0.895 0.995 0.065 0.702 0.640 0.301 0.266 0.592FAR-MIDAS vs. FAR 0.949 0.896 0.153 0.903 0.691 0.417 0.408 0.740FAR-MIDAS SV vs. FAR SV 0.951 0.917 0.156 0.882 0.659 0.489 0.501 0.579
H = 6
AR-MIDAS vs. AR 0.922 0.991 0.441 0.849 0.635 0.653 0.337 0.820AR-MIDAS SV vs. AR SV 0.898 0.970 0.278 0.796 0.632 0.960 0.087 0.672FAR-MIDAS vs. FAR 0.826 0.982 0.393 0.896 0.745 0.976 0.328 0.902FAR-MIDAS SV vs. FAR SV 0.895 0.977 0.354 0.772 0.645 0.973 0.418 0.774
H = 9
AR-MIDAS vs. AR 0.898 0.913 0.580 0.935 0.546 0.969 0.253 0.897AR-MIDAS SV vs. AR SV 0.894 0.903 0.427 0.701 0.667 0.992 0.135 0.862FAR-MIDAS vs. FAR 0.893 0.973 0.595 0.975 0.754 0.984 0.609 0.762FAR-MIDAS SV vs. FAR SV 0.931 0.989 0.633 0.795 0.869 0.990 0.258 0.730
H = 12
AR-MIDAS vs. AR 0.900 0.675 0.535 0.996 0.451 0.976 0.586 0.961AR-MIDAS SV vs. AR SV 0.880 0.406 0.850 0.968 0.559 0.953 0.827 0.936FAR-MIDAS vs. FAR 0.967 0.937 0.997 0.993 0.627 0.971 0.952 0.891FAR-MIDAS SV vs. FAR SV 0.991 0.655 0.999 0.980 0.706 0.976 0.927 0.850
Panel B: Inflation rate
H = 1
AR-MIDAS vs. AR 0.208 0.014 0.702 0.751 0.826 0.988 0.524 0.951AR-MIDAS SV vs. AR SV 0.320 0.187 0.604 0.802 0.736 0.993 0.602 0.921FAR-MIDAS vs. FAR 0.639 0.208 0.418 0.736 0.756 0.963 0.341 0.487FAR-MIDAS SV vs. FAR SV 0.780 0.853 0.448 0.695 0.708 0.983 0.568 0.408
H = 3
AR-MIDAS vs. AR 0.827 0.210 0.865 0.897 0.724 0.961 0.615 0.903AR-MIDAS SV vs. AR SV 0.915 0.448 0.770 0.401 0.511 0.730 0.541 0.838FAR-MIDAS vs. FAR 0.894 0.348 0.788 0.843 0.465 0.969 0.190 0.279FAR-MIDAS SV vs. FAR SV 0.970 0.327 0.631 0.354 0.529 0.663 0.117 0.225
H = 6
AR-MIDAS vs. AR 0.281 0.765 0.748 0.799 0.754 0.947 0.698 0.817AR-MIDAS SV vs. AR SV 0.041 0.917 0.952 0.615 0.719 0.972 0.961 0.293FAR-MIDAS vs. FAR 0.619 0.828 0.802 0.964 0.661 0.946 0.948 0.844FAR-MIDAS SV vs. FAR SV 0.470 0.818 0.895 0.183 0.509 0.970 0.728 0.057
H = 9
AR-MIDAS vs. AR 0.524 0.918 0.722 0.952 0.742 0.924 0.712 0.791AR-MIDAS SV vs. AR SV 0.582 0.997 0.341 0.225 0.287 0.323 0.480 0.302FAR-MIDAS vs. FAR 0.675 0.950 0.765 0.989 0.609 0.934 0.930 0.875FAR-MIDAS SV vs. FAR SV 0.234 0.387 0.246 0.086 0.140 0.196 0.252 0.116
H = 12
AR-MIDAS vs. AR 0.163 0.806 0.972 0.935 0.889 0.851 0.748 0.963AR-MIDAS SV vs. AR SV 0.219 0.675 0.432 0.370 0.770 0.193 0.782 0.804FAR-MIDAS vs. FAR 0.864 0.928 0.877 0.988 0.832 0.880 1.000 0.689FAR-MIDAS SV vs. FAR SV 0.886 0.850 0.736 0.303 0.752 0.367 0.946 0.262
This table reports Diebold-Mariano p-values under the null that a given MIDAS-in-mean model has the same predictive
ability as its nested non-MIDAS model, using mean squared error loss. The notation ‘AR’ refers to an autoregressive model,
‘AR-MIDAS’ refers to an augmented distributed lag MIDAS model, ‘FAR’ refers to a factor augmented autoregressive model,
and ‘FAR-MIDAS’ refers to a factor augmented distributed lag MIDAS model. The suffixes ‘SV’ and ‘SV-MIDAS’ denote
models with stochastic volatility and MIDAS stochastic volatility, respectively. For MIDAS models, the column headers
denote the daily predictor used in the regressions, namely: the effective Federal Funds rate (Ffr), the interest rate spread
between the 10-year government bond rate and the federal funds rate (Spr), value-weighted stock returns (Ret), the SML
portfolio return (Smb), the HML portfolio return (Hml), the MOM portfolio return (Mom), the business cycle variable of
Aruoba et al. (2009) (ADS), and the default spread between BAA and AAA rated bonds. The underlying p-values are
based on one-sided t-test, where the t-statistics computed with a serial correlation robust variance, using the pre-whitened
quadratic spectral estimation of Andrews and Monahan (1992). Bold numbers indicate significance at the 10 percent level.
Table 7. Out-of-sample forecast performance for IPI growth rate - Log score differentials (LSD)
ModelNo
Ffr Spr Ret Smb Hml Mom ADS DefMidas
H = 1
AR-MIDAS 0.248 0.251 0.239 0.276 0.255 0.244 0.246 0.371 0.258FAR-MIDAS 0.345 0.342 0.344 0.350 0.343 0.341 0.341 0.397 0.347AR-MIDAS SV 0.339 0.329 0.310 0.352 0.341 0.326 0.332 0.435 0.343AR-MIDAS SV-MIDAS 0.342 0.332 0.371 0.360 0.344 0.350 0.448 0.355FAR-MIDAS SV 0.386 0.380 0.387 0.384 0.379 0.376 0.374 0.443 0.380FAR-MIDAS SV-MIDAS 0.412 0.422 0.420 0.415 0.411 0.412 0.465 0.415
H = 3
AR-MIDAS 0.235 0.233 0.202 0.274 0.234 0.230 0.245 0.240 0.227FAR-MIDAS 0.281 0.271 0.272 0.291 0.272 0.276 0.277 0.277 0.270AR-MIDAS SV 0.304 0.289 0.248 0.302 0.292 0.283 0.290 0.310 0.285AR-MIDAS SV-MIDAS 0.316 0.288 0.345 0.325 0.315 0.333 0.340 0.321FAR-MIDAS SV 0.296 0.284 0.280 0.295 0.277 0.275 0.275 0.281 0.274FAR-MIDAS SV-MIDAS 0.321 0.332 0.343 0.324 0.318 0.325 0.331 0.311
H = 6
AR-MIDAS 0.291 0.291 0.273 0.302 0.284 0.286 0.283 0.300 0.279FAR-MIDAS 0.322 0.321 0.305 0.323 0.317 0.316 0.307 0.327 0.309AR-MIDAS SV 0.339 0.312 0.307 0.337 0.322 0.296 0.284 0.351 0.292AR-MIDAS SV-MIDAS 0.366 0.354 0.380 0.371 0.348 0.344 0.390 0.335FAR-MIDAS SV 0.368 0.349 0.353 0.356 0.355 0.331 0.328 0.368 0.321FAR-MIDAS SV-MIDAS 0.390 0.391 0.393 0.395 0.379 0.365 0.404 0.334
H = 9
AR-MIDAS 0.302 0.306 0.294 0.304 0.294 0.298 0.287 0.308 0.289FAR-MIDAS 0.319 0.317 0.302 0.316 0.311 0.313 0.306 0.318 0.314AR-MIDAS SV 0.337 0.328 0.304 0.319 0.314 0.299 0.288 0.346 0.308AR-MIDAS SV-MIDAS 0.386 0.337 0.371 0.364 0.359 0.342 0.384 0.345FAR-MIDAS SV 0.338 0.334 0.302 0.327 0.325 0.304 0.310 0.343 0.315FAR-MIDAS SV-MIDAS 0.394 0.351 0.380 0.375 0.368 0.357 0.387 0.368
H = 12
AR-MIDAS 0.348 0.342 0.350 0.344 0.342 0.348 0.324 0.344 0.323FAR-MIDAS 0.369 0.359 0.348 0.349 0.359 0.366 0.355 0.363 0.360AR-MIDAS SV 0.360 0.321 0.352 0.347 0.335 0.329 0.318 0.350 0.332AR-MIDAS SV-MIDAS 0.403 0.414 0.414 0.402 0.411 0.378 0.414 0.379FAR-MIDAS SV 0.361 0.325 0.343 0.329 0.338 0.328 0.321 0.351 0.347FAR-MIDAS SV-MIDAS 0.406 0.406 0.394 0.410 0.408 0.394 0.416 0.406
This table reports the average log-score (LS) differential, LSDi =∑tτ=t
(LSi,τ − LSRW,τ
), where LSi,τ (LSRW,τ ) denotes
the log-score of model i (RW), computed at time τ , and i denotes any of the models described in section 3. Positive
values of LSDi indicate that model i produces more accurate density forecasts than the RW model. The notation ‘AR’
refers to an autoregressive model, ‘AR-MIDAS’ refers to an augmented distributed lag MIDAS model, ‘FAR’ refers to a
factor augmented autoregressive model, and ‘FAR-MIDAS’ refers to a factor augmented distributed lag MIDAS model.
The suffixes ‘SV’ and ‘SV-MIDAS’ denote models with stochastic volatility and MIDAS stochastic volatility, respectively.
For MIDAS models, the column headers denote the daily predictor used in the regressions, namely: the effective Federal
Funds rate (Ffr), the interest rate spread between the 10-year government bond rate and the federal funds rate (Spr),
value-weighted stock returns (Ret), the SML portfolio return (Smb), the HML portfolio return (Hml), the MOM portfolio
return (Mom), the business cycle variable of Aruoba et al. (2009) (ADS), and the default spread between BAA and AAA
rated bonds. All forecasts and forecast errors are produced with recursive estimates of the models. The out-of-sample
period starts in 1982:1 and ends in 2011:12. Bold numbers indicate the largest LSD across all predictors for a given model.
Table 8. Out-of-sample forecast performance for inflation rate - Log score differentials (LSD)
ModelNo
Ffr Spr Ret Smb Hml Mom ADS DefMidas
H = 1
AR-MIDAS 0.141 0.159 0.158 0.136 0.140 0.134 0.139 0.139 0.137FAR-MIDAS 0.164 0.161 0.162 0.165 0.167 0.158 0.159 0.167 0.162AR-MIDAS SV 0.217 0.211 0.214 0.214 0.216 0.210 0.211 0.213 0.223AR-MIDAS SV-MIDAS 0.242 0.254 0.239 0.241 0.225 0.230 0.235 0.241FAR-MIDAS SV 0.221 0.211 0.217 0.215 0.222 0.211 0.217 0.228 0.236FAR-MIDAS SV-MIDAS 0.248 0.259 0.255 0.259 0.238 0.247 0.259 0.264
H = 3
AR-MIDAS 0.206 0.212 0.218 0.198 0.205 0.203 0.197 0.208 0.193FAR-MIDAS 0.213 0.212 0.213 0.209 0.211 0.210 0.207 0.235 0.220AR-MIDAS SV 0.267 0.254 0.261 0.258 0.269 0.258 0.264 0.257 0.262AR-MIDAS SV-MIDAS 0.298 0.322 0.302 0.303 0.302 0.306 0.294 0.291FAR-MIDAS SV 0.262 0.251 0.251 0.244 0.259 0.250 0.257 0.268 0.272FAR-MIDAS SV-MIDAS 0.299 0.307 0.295 0.297 0.294 0.304 0.313 0.316
H = 6
AR-MIDAS 0.202 0.199 0.193 0.195 0.195 0.198 0.191 0.191 0.181FAR-MIDAS 0.198 0.196 0.192 0.190 0.188 0.193 0.187 0.184 0.185AR-MIDAS SV 0.199 0.195 0.177 0.159 0.195 0.184 0.167 0.172 0.198AR-MIDAS SV-MIDAS 0.276 0.268 0.267 0.274 0.272 0.257 0.251 0.255FAR-MIDAS SV 0.189 0.184 0.166 0.148 0.204 0.183 0.146 0.179 0.193FAR-MIDAS SV-MIDAS 0.274 0.268 0.255 0.273 0.270 0.253 0.245 0.263
H = 9
AR-MIDAS 0.168 0.151 0.148 0.160 0.160 0.161 0.150 0.159 0.143FAR-MIDAS 0.170 0.156 0.157 0.160 0.161 0.165 0.156 0.148 0.150AR-MIDAS SV 0.212 0.208 0.156 0.207 0.214 0.205 0.215 0.194 0.203AR-MIDAS SV-MIDAS 0.240 0.205 0.248 0.251 0.250 0.246 0.223 0.225FAR-MIDAS SV 0.182 0.196 0.185 0.190 0.199 0.187 0.199 0.172 0.188FAR-MIDAS SV-MIDAS 0.238 0.215 0.235 0.245 0.232 0.231 0.217 0.233
H = 12
AR-MIDAS 0.212 0.219 0.208 0.206 0.204 0.199 0.203 0.208 0.193FAR-MIDAS 0.225 0.217 0.223 0.219 0.216 0.219 0.217 0.212 0.209AR-MIDAS SV 0.281 0.277 0.245 0.270 0.273 0.265 0.278 0.255 0.279AR-MIDAS SV-MIDAS 0.315 0.268 0.308 0.310 0.303 0.316 0.286 0.279FAR-MIDAS SV 0.270 0.267 0.260 0.250 0.258 0.258 0.268 0.254 0.268FAR-MIDAS SV-MIDAS 0.306 0.289 0.296 0.303 0.295 0.310 0.291 0.302
This table reports the average log-score (LS) differential, LSDi =∑tτ=t
(LSi,τ − LSRW,τ
), where LSi,τ (LSRW,τ ) denotes
the log-score of model i (RW), computed at time τ , and i denotes any of the models described in section 3. Positive
values of LSDi indicate that model i produces more accurate density forecasts than the RW model. The notation ‘AR’
refers to an autoregressive model, ‘AR-MIDAS’ refers to an augmented distributed lag MIDAS model, ‘FAR’ refers to a
factor augmented autoregressive model, and ‘FAR-MIDAS’ refers to a factor augmented distributed lag MIDAS model.
The suffixes ‘SV’ and ‘SV-MIDAS’ denote models with stochastic volatility and MIDAS stochastic volatility, respectively.
For MIDAS models, the column headers denote the daily predictor used in the regressions, namely: the effective Federal
Funds rate (Ffr), the interest rate spread between the 10-year government bond rate and the federal funds rate (Spr),
value-weighted stock returns (Ret), the SML portfolio return (Smb), the HML portfolio return (Hml), the MOM portfolio
return (Mom), the business cycle variable of Aruoba et al. (2009) (ADS), and the default spread between BAA and AAA
rated bonds. All forecasts and forecast errors are produced with recursive estimates of the models. The out-of-sample
period starts in 1982:1 and ends in 2011:12. Bold numbers indicate the largest LSD across all predictors for a given model.
Table 9. Diebold-Mariano p-values from tests of equal predictive ability - Log score differentials
Comparison Ffr Spr Ret Smb Hml Mom ADS Def
Panel A: IPI growth rate
H = 1
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV 0.156 0.065 0.102 0.095 0.089 0.093 0.175 0.164FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV 0.007 0.004 0.008 0.004 0.013 0.002 0.052 0.008
H = 3
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV 0.060 0.010 0.019 0.022 0.021 0.012 0.024 0.014FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV 0.008 0.001 0.004 0.003 0.005 0.002 0.001 0.020
H = 6
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV 0.003 0.003 0.002 0.006 0.002 0.016 0.016 0.024FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV 0.005 0.005 0.005 0.009 0.001 0.027 0.011 0.191
H = 9
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV 0.004 0.067 0.003 0.016 0.009 0.026 0.056 0.069FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV 0.003 0.015 0.002 0.009 0.003 0.017 0.016 0.011
H = 12
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV 0.003 0.006 0.004 0.005 0.001 0.026 0.009 0.072FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV 0.004 0.007 0.005 0.005 0.003 0.008 0.012 0.023
Panel B: Inflation rate
H = 1
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV 0.005 0.001 0.016 0.028 0.103 0.079 0.043 0.043FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV 0.001 0.000 0.000 0.002 0.015 0.014 0.009 0.008
H = 3
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV 0.004 0.000 0.004 0.019 0.004 0.005 0.022 0.031FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV 0.002 0.001 0.001 0.013 0.003 0.002 0.005 0.004
H = 6
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV 0.015 0.007 0.009 0.015 0.012 0.014 0.009 0.035FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV 0.014 0.006 0.008 0.023 0.012 0.009 0.019 0.022
H = 9
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV 0.085 0.035 0.017 0.039 0.017 0.055 0.092 0.108FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV 0.018 0.072 0.009 0.018 0.019 0.049 0.023 0.011
H = 12
AR-MIDAS SV-MIDAS vs. AR-MIDAS SV 0.019 0.073 0.022 0.015 0.021 0.016 0.064 0.458FAR-MIDAS SV-MIDAS vs. FAR-MIDAS SV 0.014 0.061 0.009 0.004 0.025 0.011 0.032 0.040
This table reports Diebold-Mariano p-values under the null that a given (F)AR-MIDAS SV-MIDAS model has the same
predictive ability as the (F)AR-MIDAS SV model. P-values are computed for the LSD statistic. The notation ‘AR’ refers
to an autoregressive model, ‘AR-MIDAS’ refers to an augmented distributed lag MIDAS model, ‘FAR’ refers to a factor
augmented autoregressive model, and ‘FAR-MIDAS’ refers to a factor augmented distributed lag MIDAS model. The suffixes
‘SV’ and ‘SV-MIDAS’ denote models with stochastic volatility and MIDAS stochastic volatility, respectively. For MIDAS
models, the column headers denote the daily predictor used in the regressions, namely: the effective Federal Funds rate
(Ffr), the interest rate spread between the 10-year government bond rate and the federal funds rate (Spr), value-weighted
stock returns (Ret), the SML portfolio return (Smb), the HML portfolio return (Hml), the MOM portfolio return (Mom),
the business cycle variable of Aruoba et al. (2009) (ADS), and the default spread between BAA and AAA rated bonds. The
underlying p-values are based on one-sided t-test, where the t-statistics computed with a serial correlation robust variance,
using the pre-whitened quadratic spectral estimation of Andrews and Monahan (1992). Bold numbers indicate significance
at the 10 percent level.
44
Table 10. Out-of-sample forecast performance - model combinations
IPI growth rate
ModelPanel A: RMSFE Panel B: LSD
H = 1 H = 3 H = 6 H = 9 H = 12 H = 1 H = 3 H = 6 H = 9 H = 12
Equal-weighted combination 0.675 0.780 0.766 0.743 0.697 0.407 0.338 0.386 0.379 0.419Bayesian model averaging 0.649 0.787 0.770 0.761 0.723 0.453 0.325 0.382 0.369 0.366Optimal prediction pool 0.657 0.780 0.774 0.758 0.717 0.446 0.331 0.376 0.368 0.374
Panel C: RMSFE vs. AR model Panel D: PL vs. AR model
H = 1 H = 3 H = 6 H = 9 H = 12 H = 1 H = 3 H = 6 H = 9 H = 12
Equal-weighted combination 0.000 0.016 0.089 0.321 0.214 0.000 0.000 0.000 0.001 0.005Bayesian model averaging 0.001 0.110 0.240 0.823 0.838 0.000 0.001 0.001 0.027 0.360Optimal prediction pool 0.000 0.027 0.228 0.818 0.772 0.000 0.001 0.002 0.020 0.295
Panel E: RMSFE vs. AR SV model Panel F: PL vs. AR-SV model
H = 1 H = 3 H = 6 H = 9 H = 12 H = 1 H = 3 H = 6 H = 9 H = 12
Equal-weighted combination 0.000 0.010 0.032 0.141 0.026 0.001 0.096 0.051 0.073 0.031Bayesian model averaging 0.002 0.089 0.219 0.796 0.740 0.000 0.179 0.072 0.158 0.447Optimal prediction pool 0.001 0.015 0.212 0.709 0.611 0.000 0.098 0.071 0.136 0.358
Inflation rate
ModelPanel A: RMSFE Panel B: LSD
H = 1 H = 3 H = 6 H = 9 H = 12 H = 1 H = 3 H = 6 H = 9 H = 12
Equal-weighted combination 0.908 0.812 0.778 0.797 0.759 0.229 0.283 0.254 0.227 0.286Bayesian model averaging 0.910 0.805 0.783 0.799 0.760 0.252 0.309 0.261 0.228 0.292Optimal prediction pool 0.912 0.797 0.773 0.787 0.764 0.241 0.308 0.266 0.233 0.282
Panel C: RMSFE vs. AR model Panel D: PL vs. AR model
H = 1 H = 3 H = 6 H = 9 H = 12 H = 1 H = 3 H = 6 H = 9 H = 12
Equal-weighted combination 0.033 0.347 0.247 0.175 0.004 0.000 0.000 0.009 0.003 0.000Bayesian model averaging 0.129 0.096 0.375 0.239 0.011 0.000 0.000 0.022 0.019 0.001Optimal prediction pool 0.182 0.049 0.193 0.066 0.036 0.000 0.000 0.012 0.020 0.006
Panel E: RMSFE vs. AR SV model Panel F: PL vs. AR-SV model
H = 1 H = 3 H = 6 H = 9 H = 12 H = 1 H = 3 H = 6 H = 9 H = 12
Equal-weighted combination 0.428 0.427 0.141 0.450 0.868 0.154 0.145 0.018 0.222 0.378Bayesian model averaging 0.528 0.147 0.167 0.477 0.870 0.004 0.011 0.012 0.214 0.294Optimal prediction pool 0.621 0.061 0.061 0.252 0.927 0.035 0.009 0.005 0.112 0.463
This table reports out-of-sample results for the optimal predictive pool of Geweke and Amisano (2011), an equal-weighted
model combination scheme, and Bayesian Model Averaging applied to 36 MIDAS forecasting models that use different daily
predictors and make different choices regarding the inclusion or exclusion of additional macro factors and the volatility
dynamics of US quarterly GDP growth. In each case the models and combination weights are estimated recursively using
only data up to the point of the forecast. Panels A and B report various measures of point and density forecast performance
over two out-of-sample periods, namely 1982:1 to 2011:12 (panel A), and 2001:I to 2008:IV (panel B). The statistics
reported are: RMSFEi, the ratio between the RMSFE of model combination i and the RMSFE of the Random Walk (RW)
model; LSDi, the average log-score differential between model combination i and the RW model; CRPSDi, the average
continuously ranked probability score differential of model combination i relative to the RW model. Values less than one
for RMSFEi indicate that model combination i produces more accurate point forecasts than the RW model; positive
values of LSDi indicate that model combination i produces more accurate density forecasts than the RW model; values
less than one for CRPSDi suggest that model combination i performs better than the benchmark RW model. Panels C
and D report the Diebold-Mariano p-values under the null that a given model combination has the same predictive ability
than the benchmark model, over the same two out-of-sample periods, 1982:1 to 2011:12 (panel C), and 2001:I to 2008:IV
(panel D). P-values are computed for the RMSFE, LSD, and the CRPSD statistics, both against the Random Walk (RW)
benchmark and the autoregressive (AR) benchmark, and are based on one-sided t-tests, where the t-statistics are computed
with a serial correlation robust variance, using the pre-whitened quadratic spectral estimation of Andrews and Monahan
(1992). Bold numbers in Panels C-F indicate significance at the 10 percent level.
Figure 1. Mean forecasts of MIDAS models
1977 1982 1987 1992 1997 2002 2007 2012
Mea
n fo
reca
st
-5
-4
-3
-2
-1
0
1
2
3IPI growth rate, H=1
Random WalkAR-MIDASAR-MIDAS SVAR-MIDAS SV-MIDAS
1977 1982 1987 1992 1997 2002 2007 2012
Mea
n fo
reca
st
-5
-4
-3
-2
-1
0
1
2
3IPI growth rate, H=12
1977 1982 1987 1992 1997 2002 2007 2012
Mea
n fo
reca
st
-30
-20
-10
0
10
20Inflation rate, H=1
1977 1982 1987 1992 1997 2002 2007 2012
Mea
n fo
reca
st
-30
-20
-10
0
10
20Inflation rate, H=12
This figure shows the recursive conditional mean forecasts computed using the Random Walk model (dashed back line)
as well as the predictive distributions of the AR-MIDAS (solid blue line), the AR-MIDAS SV (dashed red line), and the
AR-MIDAS SV-MIDAS (green dashed-dotted line) models. All MIDAS models displayed in the two panels use the Ffr daily
series as predictor. All models are estimated recursively over the out-of-sample period, which starts in 1982:1 and ends in
2011:12.
46
Figure 2. Volatility forecasts of MIDAS models
1977 1982 1987 1992 1997 2002 2007 2012
Vol
atili
ty fo
reca
st
0
0.5
1
1.5
2IPI growth rate, H=1
AR-MIDASAR-MIDAS SVAR-MIDAS SV-MIDAS
1977 1982 1987 1992 1997 2002 2007 2012
Vol
atili
ty fo
reca
st
0
0.5
1
1.5
2IPI growth rate, H=12
1977 1982 1987 1992 1997 2002 2007 2012
Vol
atili
ty fo
reca
st
0
2
4
6
8
10
12Inflation rate, H=1
1977 1982 1987 1992 1997 2002 2007 2012
Vol
atili
ty fo
reca
st
0
2
4
6
8
10
12Inflation rate, H=12
This figure shows the recursive conditional volatility forecasts computed using the predictive distributions of the AR-MIDAS
(solid blue line), the AR-MIDAS SV (dashed red line), and the AR-MIDAS SV-MIDAS (green dashed-dotted line) models.
All MIDAS models displayed in the two panels use the Ffr daily series as predictor. All models are estimated recursively
over the out-of-sample period, which starts in 1982:1 and ends in 2011:12.
47
Figure 3. Predictive density of IPI growth and inflation under different AR-MIDAS models
-15 -10 -5 0 5 10
Fre
q.
0
0.2
0.4
0.6
0.8
1
1.2
1.4IPI growth rate (Jan1994)
AR-MIDASAR-MIDAS SVAR-MIDAS SV-MIDAS
-15 -10 -5 0 5 10
Fre
q.
0
0.1
0.2
0.3
0.4
0.5
0.6IPI growth rate (Dec2008)
AR-MIDASAR-MIDAS SVAR-MIDAS SV-MIDAS
-80 -60 -40 -20 0 20 40 60
Fre
q.
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35Inflation rate (Jan1994)
AR-MIDASAR-MIDAS SVAR-MIDAS SV-MIDAS
-80 -60 -40 -20 0 20 40 60
Fre
q.
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14Inflation rate (Dec2008)
AR-MIDASAR-MIDAS SVAR-MIDAS SV-MIDAS
This figure shows the predictive density for the IPI growth rate (top panels) and the inflation rate (bottom panels) under
alternative MIDAS models, produced in pseudo real-time for 1994:1 (left panels) and 2008:12 (right panels). The top panels
display the predictive density for IPI growth under three alternative MIDAS models, where the blue solid line corresponds
to the AR-MIDAS model, the red dashed line refers to the AR-MIDAS SV model, and the green dashed-dotted line refers
to the AR-MIDAS SV-MIDAS model. The bottom panels display the predictive density for inflation using the same three
MIDAS models. All MIDAS models use the daily Ffr series as predictor, and all predictive densities are produced using
recursive model estimates.
48
Figure 4. Cumulative sum of log-score differentials (CLSD) for AR-MIDAS SV models
1977 1982 1987 1992 1997 2002 2007 2012
∆ c
umul
ativ
e lo
g sc
ore
-50
0
50
100
150
200IPI growth rate, H=1
AR SVAR-MIDAS SV (best)AR-MIDAS SV-MIDAS (best)
1977 1982 1987 1992 1997 2002 2007 2012
∆ c
umul
ativ
e lo
g sc
ore
-50
0
50
100
150
200IPI growth rate, H=12
AR SVAR-MIDAS SV (best)AR-MIDAS SV-MIDAS (best)
1977 1982 1987 1992 1997 2002 2007 2012
∆ c
umul
ativ
e lo
g sc
ore
-20
0
20
40
60
80
100
120Inflation rate, H=1
AR SVAR-MIDAS SV (best)AR-MIDAS SV-MIDAS (best)
1977 1982 1987 1992 1997 2002 2007 2012
∆ c
umul
ativ
e lo
g sc
ore
-20
0
20
40
60
80
100
120Inflation rate, H=12
AR SVAR-MIDAS SV (best)AR-MIDAS SV-MIDAS (best)
This figure shows the sum of log predictive scores from AR-MIDAS SV and AR SV models, computed relative the sum
of log predictive scores of the Random Walk (RW) model. Each quarter we estimate the parameters of the forecast
models recursively and generate one-step-ahead density forecasts of real GDP growth which are in turn used to compute
log-predictive scores. This procedure is applied to the benchmark RW model as well as to all the alternative forecasting
models. We then plot the cumulative sum of log-predictive scores (LSt) for the alternative models computed relative to
the cumulative sum of log-predictive scores of the RW model, LSt − LSRWt . Values above zero indicate that a forecast
model generates better performance than the RW benchmark, while negative values suggest the opposite. The top two
panels compare the forecasting performance of different autoregressive models for IPI growth with time-varying volatility
and different forecast horizons, while the bottom two panels focuses on Inflation.
49
Figure 5. Relation between RMSFE and other measures of predictive performance for IPI growthrate
Relative RMSFE0.62 0.64 0.66 0.68 0.7 0.72 0.74 0.76 0.78
LSD
0.25
0.3
0.35
0.4
0.45
RMSFE vs. LSD, H=1
Linear modelsSV modelsSV-MIDAS models
Relative RMSFE0.62 0.64 0.66 0.68 0.7 0.72 0.74 0.76 0.78
CR
PS
D
0.65
0.7
0.75
0.8RMSFE vs. CRPSD, H=1
Relative RMSFE0.67 0.68 0.69 0.7 0.71 0.72 0.73 0.74 0.75 0.76
LSD
0.3
0.32
0.34
0.36
0.38
0.4
0.42
0.44RMSFE vs. LSD, H=12
Relative RMSFE0.67 0.68 0.69 0.7 0.71 0.72 0.73 0.74 0.75 0.76
CR
PS
D
0.66
0.68
0.7
0.72
0.74
RMSFE vs. CRPSD, H=12
This figure presents scatter plots of out-of-sample predictive performance measures against the RMSFE (root mean squared
forecast error) measure. All estimates of predictive performance are measured relative to the Random Walk (RW) model,
so that the origin in each figure, shown as the intersection of the solid black lines, corresponds to the RW model. In
addition, the intersection of the black dashed lines displays the relative out-of-sample performance of the AR model. The
left panel displays the relation between RMSFE and the average log-score differential (LSD), while the right panel shows the
relation between RMSFE and the continuously ranked probability score differential (CRPSD). Blue squares correspond to
linear models (including AR, AR-MIDAS, FAR, and FAR-MIDAS models), red diamonds represent the stochastic volatility
models (including AR SV, AR-MIDAS SV, FAR SV, and FAR-MIDAS SV), and the green triangles depict the MIDAS in
volatility models (including AR-MIDAS SV-MIDAS and FAR-MIDAS SV-MIDAS). All measures of predictive performance
are produced with recursive estimates of the models. The out-of-sample period starts in 1982:1 and ends in 2011:12.
50
Figure 6. Relation between RMSFE and other measures of predictive performance for Inflationrate
Relative RMSFE0.88 0.89 0.9 0.91 0.92 0.93 0.94 0.95
LSD
0.15
0.2
0.25
RMSFE vs. LSD, H=1
Linear modelsSV modelsSV-MIDAS models
Relative RMSFE0.88 0.89 0.9 0.91 0.92 0.93 0.94 0.95
CR
PS
D0.82
0.84
0.86
0.88
0.9
0.92RMSFE vs. CRPSD, H=1
Relative RMSFE0.72 0.74 0.76 0.78 0.8 0.82
LSD
0.2
0.25
0.3
RMSFE vs. LSD, H=12
Relative RMSFE0.72 0.74 0.76 0.78 0.8 0.82
CR
PS
D
0.72
0.74
0.76
0.78
0.8
0.82
0.84RMSFE vs. CRPSD, H=12
This figure presents scatter plots of out-of-sample predictive performance measures against the RMSFE (root mean squared
forecast error) measure. All estimates of predictive performance are measured relative to the Random Walk (RW) model,
so that the origin in each figure, shown as the intersection of the solid black lines, corresponds to the RW model. In
addition, the intersection of the black dashed lines displays the relative out-of-sample performance of the AR model. The
left panel displays the relation between RMSFE and the average log-score differential (LSD), while the right panel shows the
relation between RMSFE and the continuously ranked probability score differential (CRPSD). Blue squares correspond to
linear models (including AR, AR-MIDAS, FAR, and FAR-MIDAS models), red diamonds represent the stochastic volatility
models (including AR SV, AR-MIDAS SV, FAR SV, and FAR-MIDAS SV), and the green triangles depict the MIDAS in
volatility models (including AR-MIDAS SV-MIDAS and FAR-MIDAS SV-MIDAS). All measures of predictive performance
are produced with recursive estimates of the models. The out-of-sample period starts in 1982:1 and ends in 2011:12.
51
Figure 7. Weights on different model classes in the optimal prediction pool
IPI growth rate, H=1
1982 1987 1992 1997 2002 2007
Opt
imal
pre
dict
ion
pool
wei
ghts
0
0.2
0.4
0.6
0.8
1AR-MIDASAR-MIDAS SVAR-MIDAS SV-MIDASFAR-MIDASFAR-MIDAS SVFAR-MIDAS SV-MIDAS
IPI growth rate, H=12
1982 1987 1992 1997 2002 2007O
ptim
al p
redi
ctio
n po
ol w
eigh
ts0
0.2
0.4
0.6
0.8
1
Inflation rate, H=1
1982 1987 1992 1997 2002 2007
Opt
imal
pre
dict
ion
pool
wei
ghts
0
0.2
0.4
0.6
0.8
1Inflation rate, H=12
1982 1987 1992 1997 2002 2007
Opt
imal
pre
dict
ion
pool
wei
ghts
0
0.2
0.4
0.6
0.8
1
This figure plots the optimal weights on different models in the predictive pool, computed in real time by solving theminimization problem
w∗t = arg maxw
t−1∑τ=1
log
[N∑i=1
wi × Sτ+1,i
]
where N = 36 is the number of models considered, and the solution is found subject to w∗t belonging to the N−dimensional
unit simplex. Sτ+1,i denotes the time τ + 1 recursively computed log score for model i, i.e. Sτ+1,i = exp (LSτ+1,i). Dark
blue bars show the weights on the AR-MIDAS models in the optimal prediction pool, blue bars show the weights assigned
to the AR-MIDAS SV models, and light blue bars show the weights assigned to the AR-MIDAS SV-MIDAS models; gray
bars show the weights on the FAR-MIDAS models in the optimal prediction pool, orange bars show the weights assigned
to the FAR-MIDAS SV models, and yellow bars show the weights assigned to the FAR-MIDAS SV-MIDAS models.
52
Figure 8. Weights on individual daily predictors in the optimal prediction pool
IPI growth rate, H=1
1982 1987 1992 1997 2002 2007
Opt
imal
pre
dict
ion
pool
wei
ghts
0
0.2
0.4
0.6
0.8
1FfrSprRetSmbHmlMomADSDef
IPI growth rate, H=12
1982 1987 1992 1997 2002 2007
Opt
imal
pre
dict
ion
pool
wei
ghts
0
0.2
0.4
0.6
0.8
1
Inflation rate, H=1
1982 1987 1992 1997 2002 2007
Opt
imal
pre
dict
ion
pool
wei
ghts
0
0.2
0.4
0.6
0.8
1Inflation rate, H=12
1982 1987 1992 1997 2002 2007
Opt
imal
pre
dict
ion
pool
wei
ghts
0
0.2
0.4
0.6
0.8
1
This figure plots the optimal weights on different daily predictors in the predictive pool, computed in real time by solvingthe minimization problem
w∗t = arg maxw
t−1∑τ=1
log
[N∑i=1
wi × Sτ+1,i
]
where N = 36 is the number of models considered, and the solution is found subject to w∗t belonging to the N−dimensional
unit simplex. Sτ+1,i denotes the time τ + 1 recursively computed log score for model i, i.e. Sτ+1,i = exp (LSτ+1,i). Dark
blue bars show the weights associated with the effective Federal Funds rate (Ffr) predictor in the optimal prediction pool,
blue bars show the weights associated with the 10-year government bond rate and the federal fund rate (Spr), and light
blue bars show the weights assigned to the value-weighted stock returns (Ret); celeste bars show the weights associated
with the SML portfolio return (Smb), and green bars show the weights assigned to he HML portfolio return (Hml); gray,
orange, and yellow bars show the weights associated with the MOM portfolio return (Mom), the Aruoba-Diebold-Scotti
index (ADS), and the Default spread (Def)
53